diff --git "a/nvidia_parakeet-v2/AudioEncoder.mlmodelc/model.mil" "b/nvidia_parakeet-v2/AudioEncoder.mlmodelc/model.mil"
--- "a/nvidia_parakeet-v2/AudioEncoder.mlmodelc/model.mil"
+++ "b/nvidia_parakeet-v2/AudioEncoder.mlmodelc/model.mil"
@@ -1,14 +1,17 @@
 program(1.0)
 [buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3401.3.1"}, {"coremlc-version", "3401.4.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.2"}})]
 {
-    func main<ios17>(tensor<fp16, [1, 1, 1501, 128]> melspectrogram_features) {
+    func main<ios17>(tensor<fp16, [1, 1, 1, 1]> input_1, tensor<fp16, [1, 1, 1501, 128]> melspectrogram_features) {
+            tensor<fp16, [1, 1024, 1, 375]> pos_emb_to_fp16 = const()[name = tensor<string, []>("pos_emb_to_fp16"), val = tensor<fp16, [1, 1024, 1, 375]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp16, [1, 1024, 1, 375]> var_55_cast_fp16 = sub(x = pos_emb_to_fp16, y = input_1)[name = tensor<string, []>("op_55_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 375]> obj_3_cast_fp16 = mul(x = var_55_cast_fp16, y = input_1)[name = tensor<string, []>("obj_3_cast_fp16")];
             tensor<string, []> input_1_pad_type_0 = const()[name = tensor<string, []>("input_1_pad_type_0"), val = tensor<string, []>("custom")];
             tensor<int32, [4]> input_1_pad_0 = const()[name = tensor<string, []>("input_1_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
             tensor<int32, [2]> input_1_strides_0 = const()[name = tensor<string, []>("input_1_strides_0"), val = tensor<int32, [2]>([2, 2])];
             tensor<int32, [2]> input_1_dilations_0 = const()[name = tensor<string, []>("input_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> input_1_groups_0 = const()[name = tensor<string, []>("input_1_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [256, 1, 3, 3]> pre_encode_conv_0_weight_to_fp16 = const()[name = tensor<string, []>("pre_encode_conv_0_weight_to_fp16"), val = tensor<fp16, [256, 1, 3, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
-            tensor<fp16, [256]> pre_encode_conv_0_bias_to_fp16 = const()[name = tensor<string, []>("pre_encode_conv_0_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4736)))];
+            tensor<fp16, [256, 1, 3, 3]> pre_encode_conv_0_weight_to_fp16 = const()[name = tensor<string, []>("pre_encode_conv_0_weight_to_fp16"), val = tensor<fp16, [256, 1, 3, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(768128)))];
+            tensor<fp16, [256]> pre_encode_conv_0_bias_to_fp16 = const()[name = tensor<string, []>("pre_encode_conv_0_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(772800)))];
             tensor<fp16, [1, 256, 751, 64]> input_1_cast_fp16 = conv(bias = pre_encode_conv_0_bias_to_fp16, dilations = input_1_dilations_0, groups = input_1_groups_0, pad = input_1_pad_0, pad_type = input_1_pad_type_0, strides = input_1_strides_0, weight = pre_encode_conv_0_weight_to_fp16, x = melspectrogram_features)[name = tensor<string, []>("input_1_cast_fp16")];
             tensor<fp16, [1, 256, 751, 64]> input_3_cast_fp16 = relu(x = input_1_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
             tensor<string, []> input_5_pad_type_0 = const()[name = tensor<string, []>("input_5_pad_type_0"), val = tensor<string, []>("custom")];
@@ -16,16 +19,16 @@ program(1.0)
             tensor<int32, [2]> input_5_strides_0 = const()[name = tensor<string, []>("input_5_strides_0"), val = tensor<int32, [2]>([2, 2])];
             tensor<int32, []> input_5_groups_0 = const()[name = tensor<string, []>("input_5_groups_0"), val = tensor<int32, []>(256)];
             tensor<int32, [2]> input_5_dilations_0 = const()[name = tensor<string, []>("input_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<fp16, [256, 1, 3, 3]> pre_encode_conv_2_weight_to_fp16 = const()[name = tensor<string, []>("pre_encode_conv_2_weight_to_fp16"), val = tensor<fp16, [256, 1, 3, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5312)))];
-            tensor<fp16, [256]> pre_encode_conv_2_bias_to_fp16 = const()[name = tensor<string, []>("pre_encode_conv_2_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9984)))];
+            tensor<fp16, [256, 1, 3, 3]> pre_encode_conv_2_weight_to_fp16 = const()[name = tensor<string, []>("pre_encode_conv_2_weight_to_fp16"), val = tensor<fp16, [256, 1, 3, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(773376)))];
+            tensor<fp16, [256]> pre_encode_conv_2_bias_to_fp16 = const()[name = tensor<string, []>("pre_encode_conv_2_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(778048)))];
             tensor<fp16, [1, 256, 376, 32]> input_5_cast_fp16 = conv(bias = pre_encode_conv_2_bias_to_fp16, dilations = input_5_dilations_0, groups = input_5_groups_0, pad = input_5_pad_0, pad_type = input_5_pad_type_0, strides = input_5_strides_0, weight = pre_encode_conv_2_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
             tensor<string, []> input_7_pad_type_0 = const()[name = tensor<string, []>("input_7_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> input_7_strides_0 = const()[name = tensor<string, []>("input_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> input_7_pad_0 = const()[name = tensor<string, []>("input_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> input_7_dilations_0 = const()[name = tensor<string, []>("input_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> input_7_groups_0 = const()[name = tensor<string, []>("input_7_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [256, 256, 1, 1]> pre_encode_conv_3_weight_to_fp16 = const()[name = tensor<string, []>("pre_encode_conv_3_weight_to_fp16"), val = tensor<fp16, [256, 256, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10560)))];
-            tensor<fp16, [256]> pre_encode_conv_3_bias_to_fp16 = const()[name = tensor<string, []>("pre_encode_conv_3_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141696)))];
+            tensor<fp16, [256, 256, 1, 1]> pre_encode_conv_3_weight_to_fp16 = const()[name = tensor<string, []>("pre_encode_conv_3_weight_to_fp16"), val = tensor<fp16, [256, 256, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(778624)))];
+            tensor<fp16, [256]> pre_encode_conv_3_bias_to_fp16 = const()[name = tensor<string, []>("pre_encode_conv_3_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(909760)))];
             tensor<fp16, [1, 256, 376, 32]> input_7_cast_fp16 = conv(bias = pre_encode_conv_3_bias_to_fp16, dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = pre_encode_conv_3_weight_to_fp16, x = input_5_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
             tensor<fp16, [1, 256, 376, 32]> input_9_cast_fp16 = relu(x = input_7_cast_fp16)[name = tensor<string, []>("input_9_cast_fp16")];
             tensor<string, []> input_11_pad_type_0 = const()[name = tensor<string, []>("input_11_pad_type_0"), val = tensor<string, []>("custom")];
@@ -33,252 +36,251 @@ program(1.0)
             tensor<int32, [2]> input_11_strides_0 = const()[name = tensor<string, []>("input_11_strides_0"), val = tensor<int32, [2]>([2, 2])];
             tensor<int32, []> input_11_groups_0 = const()[name = tensor<string, []>("input_11_groups_0"), val = tensor<int32, []>(256)];
             tensor<int32, [2]> input_11_dilations_0 = const()[name = tensor<string, []>("input_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<fp16, [256, 1, 3, 3]> pre_encode_conv_5_weight_to_fp16 = const()[name = tensor<string, []>("pre_encode_conv_5_weight_to_fp16"), val = tensor<fp16, [256, 1, 3, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142272)))];
-            tensor<fp16, [256]> pre_encode_conv_5_bias_to_fp16 = const()[name = tensor<string, []>("pre_encode_conv_5_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146944)))];
+            tensor<fp16, [256, 1, 3, 3]> pre_encode_conv_5_weight_to_fp16 = const()[name = tensor<string, []>("pre_encode_conv_5_weight_to_fp16"), val = tensor<fp16, [256, 1, 3, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(910336)))];
+            tensor<fp16, [256]> pre_encode_conv_5_bias_to_fp16 = const()[name = tensor<string, []>("pre_encode_conv_5_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(915008)))];
             tensor<fp16, [1, 256, 188, 16]> input_11_cast_fp16 = conv(bias = pre_encode_conv_5_bias_to_fp16, dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = pre_encode_conv_5_weight_to_fp16, x = input_9_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
             tensor<string, []> input_13_pad_type_0 = const()[name = tensor<string, []>("input_13_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> input_13_strides_0 = const()[name = tensor<string, []>("input_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> input_13_pad_0 = const()[name = tensor<string, []>("input_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> input_13_dilations_0 = const()[name = tensor<string, []>("input_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> input_13_groups_0 = const()[name = tensor<string, []>("input_13_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [256, 256, 1, 1]> pre_encode_conv_6_weight_to_fp16 = const()[name = tensor<string, []>("pre_encode_conv_6_weight_to_fp16"), val = tensor<fp16, [256, 256, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(147520)))];
-            tensor<fp16, [256]> pre_encode_conv_6_bias_to_fp16 = const()[name = tensor<string, []>("pre_encode_conv_6_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278656)))];
+            tensor<fp16, [256, 256, 1, 1]> pre_encode_conv_6_weight_to_fp16 = const()[name = tensor<string, []>("pre_encode_conv_6_weight_to_fp16"), val = tensor<fp16, [256, 256, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(915584)))];
+            tensor<fp16, [256]> pre_encode_conv_6_bias_to_fp16 = const()[name = tensor<string, []>("pre_encode_conv_6_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1046720)))];
             tensor<fp16, [1, 256, 188, 16]> input_13_cast_fp16 = conv(bias = pre_encode_conv_6_bias_to_fp16, dilations = input_13_dilations_0, groups = input_13_groups_0, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = input_13_strides_0, weight = pre_encode_conv_6_weight_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")];
             tensor<fp16, [1, 256, 188, 16]> x_1_cast_fp16 = relu(x = input_13_cast_fp16)[name = tensor<string, []>("x_1_cast_fp16")];
             tensor<int32, [4]> var_115_perm_0 = const()[name = tensor<string, []>("op_115_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
             tensor<int32, [4]> var_118 = const()[name = tensor<string, []>("op_118"), val = tensor<int32, [4]>([1, 4096, 1, 188])];
             tensor<fp16, [1, 256, 16, 188]> var_115_cast_fp16 = transpose(perm = var_115_perm_0, x = x_1_cast_fp16)[name = tensor<string, []>("transpose_0")];
-            tensor<fp16, [1, 4096, 1, 188]> var_119_cast_fp16 = reshape(shape = var_118, x = var_115_cast_fp16)[name = tensor<string, []>("op_119_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_15_cast_fp16 = reshape(shape = var_118, x = var_115_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
             tensor<string, []> inputs_1_pad_type_0 = const()[name = tensor<string, []>("inputs_1_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> inputs_1_strides_0 = const()[name = tensor<string, []>("inputs_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> inputs_1_pad_0 = const()[name = tensor<string, []>("inputs_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> inputs_1_dilations_0 = const()[name = tensor<string, []>("inputs_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> inputs_1_groups_0 = const()[name = tensor<string, []>("inputs_1_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_123_to_fp16 = const()[name = tensor<string, []>("op_123_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(279232)))];
-            tensor<fp16, [1024]> pre_encode_out_bias_to_fp16 = const()[name = tensor<string, []>("pre_encode_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8667904)))];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_1_cast_fp16 = conv(bias = pre_encode_out_bias_to_fp16, dilations = inputs_1_dilations_0, groups = inputs_1_groups_0, pad = inputs_1_pad_0, pad_type = inputs_1_pad_type_0, strides = inputs_1_strides_0, weight = var_123_to_fp16, x = var_119_cast_fp16)[name = tensor<string, []>("inputs_1_cast_fp16")];
-            tensor<int32, []> var_133 = const()[name = tensor<string, []>("op_133"), val = tensor<int32, []>(3)];
+            tensor<fp16, [1024, 4096, 1, 1]> pre_encode_out_weight_to_fp16 = const()[name = tensor<string, []>("pre_encode_out_weight_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1047296)))];
+            tensor<fp16, [1024]> pre_encode_out_bias_to_fp16 = const()[name = tensor<string, []>("pre_encode_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9435968)))];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_1_cast_fp16 = conv(bias = pre_encode_out_bias_to_fp16, dilations = inputs_1_dilations_0, groups = inputs_1_groups_0, pad = inputs_1_pad_0, pad_type = inputs_1_pad_type_0, strides = inputs_1_strides_0, weight = pre_encode_out_weight_to_fp16, x = input_15_cast_fp16)[name = tensor<string, []>("inputs_1_cast_fp16")];
+            tensor<int32, []> var_131 = const()[name = tensor<string, []>("op_131"), val = tensor<int32, []>(3)];
             tensor<int32, [1]> out_1_axes_0 = const()[name = tensor<string, []>("out_1_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_164_to_fp16 = const()[name = tensor<string, []>("op_164_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_164_to_fp16, x = inputs_1_cast_fp16)[name = tensor<string, []>("out_1_cast_fp16")];
-            tensor<fp16, [1024]> input_15_mean_0_to_fp16 = const()[name = tensor<string, []>("input_15_mean_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8670016)))];
-            tensor<fp16, [1024]> input_15_variance_0_to_fp16 = const()[name = tensor<string, []>("input_15_variance_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8672128)))];
-            tensor<fp16, [1024]> input_15_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_15_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8674240)))];
-            tensor<fp16, [1024]> input_15_beta_0_to_fp16 = const()[name = tensor<string, []>("input_15_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8676352)))];
-            tensor<fp16, []> input_15_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_15_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_15_cast_fp16 = batch_norm(beta = input_15_beta_0_to_fp16, epsilon = input_15_epsilon_0_to_fp16, gamma = input_15_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_1_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
-            tensor<string, []> input_17_pad_type_0 = const()[name = tensor<string, []>("input_17_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_17_strides_0 = const()[name = tensor<string, []>("input_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_17_pad_0 = const()[name = tensor<string, []>("input_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_17_dilations_0 = const()[name = tensor<string, []>("input_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_17_groups_0 = const()[name = tensor<string, []>("input_17_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [4096, 1024, 1, 1]> layers_0_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_feed_forward1_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8678464)))];
-            tensor<fp16, [4096]> layers_0_feed_forward1_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_feed_forward1_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17067136)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_17_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_17_dilations_0, groups = input_17_groups_0, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = input_17_strides_0, weight = layers_0_feed_forward1_fc1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor<string, []>("input_17_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_19_cast_fp16 = silu(x = input_17_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")];
+            tensor<fp16, []> var_162_to_fp16 = const()[name = tensor<string, []>("op_162_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_162_to_fp16, x = inputs_1_cast_fp16)[name = tensor<string, []>("out_1_cast_fp16")];
+            tensor<fp16, [1024]> input_17_mean_0_to_fp16 = const()[name = tensor<string, []>("input_17_mean_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9438080)))];
+            tensor<fp16, [1024]> input_17_variance_0_to_fp16 = const()[name = tensor<string, []>("input_17_variance_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9440192)))];
+            tensor<fp16, [1024]> input_17_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_17_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9442304)))];
+            tensor<fp16, [1024]> input_17_beta_0_to_fp16 = const()[name = tensor<string, []>("input_17_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9444416)))];
+            tensor<fp16, []> input_17_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_17_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_17_cast_fp16 = batch_norm(beta = input_17_beta_0_to_fp16, epsilon = input_17_epsilon_0_to_fp16, gamma = input_17_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_1_cast_fp16)[name = tensor<string, []>("input_17_cast_fp16")];
+            tensor<string, []> input_19_pad_type_0 = const()[name = tensor<string, []>("input_19_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_19_strides_0 = const()[name = tensor<string, []>("input_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_19_pad_0 = const()[name = tensor<string, []>("input_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_19_dilations_0 = const()[name = tensor<string, []>("input_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_19_groups_0 = const()[name = tensor<string, []>("input_19_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [4096, 1024, 1, 1]> layers_0_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_feed_forward1_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9446528)))];
+            tensor<fp16, [4096]> layers_0_feed_forward1_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_feed_forward1_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17835200)))];
+            tensor<fp16, [1, 4096, 1, 188]> input_19_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_19_dilations_0, groups = input_19_groups_0, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = input_19_strides_0, weight = layers_0_feed_forward1_fc1_weight_to_fp16, x = input_17_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_21_cast_fp16 = silu(x = input_19_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")];
             tensor<string, []> x_3_pad_type_0 = const()[name = tensor<string, []>("x_3_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_3_strides_0 = const()[name = tensor<string, []>("x_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_3_pad_0 = const()[name = tensor<string, []>("x_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_3_dilations_0 = const()[name = tensor<string, []>("x_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_3_groups_0 = const()[name = tensor<string, []>("x_3_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_192_weight_0_to_fp16 = const()[name = tensor<string, []>("op_192_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17075392)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_192_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_3_dilations_0, groups = x_3_groups_0, pad = x_3_pad_0, pad_type = x_3_pad_type_0, strides = x_3_strides_0, weight = var_192_weight_0_to_fp16, x = input_19_cast_fp16)[name = tensor<string, []>("op_192_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = var_192_cast_fp16)[name = tensor<string, []>("inputs_3_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_190_weight_0_to_fp16 = const()[name = tensor<string, []>("op_190_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17843456)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_190_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_3_dilations_0, groups = x_3_groups_0, pad = x_3_pad_0, pad_type = x_3_pad_type_0, strides = x_3_strides_0, weight = var_190_weight_0_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_190_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = var_190_cast_fp16)[name = tensor<string, []>("inputs_3_cast_fp16")];
             tensor<int32, [1]> out_3_axes_0 = const()[name = tensor<string, []>("out_3_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_202_to_fp16 = const()[name = tensor<string, []>("op_202_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_202_to_fp16, x = inputs_3_cast_fp16)[name = tensor<string, []>("out_3_cast_fp16")];
-            tensor<fp16, [1024]> obj_1_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_1_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25464064)))];
-            tensor<fp16, [1024]> obj_1_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_1_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25466176)))];
+            tensor<fp16, []> var_200_to_fp16 = const()[name = tensor<string, []>("op_200_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_200_to_fp16, x = inputs_3_cast_fp16)[name = tensor<string, []>("out_3_cast_fp16")];
+            tensor<fp16, [1024]> obj_1_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_1_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26232128)))];
+            tensor<fp16, [1024]> obj_1_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_1_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26234240)))];
             tensor<fp16, []> obj_1_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_1_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> obj_1_cast_fp16 = batch_norm(beta = obj_1_beta_0_to_fp16, epsilon = obj_1_epsilon_0_to_fp16, gamma = obj_1_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_3_cast_fp16)[name = tensor<string, []>("obj_1_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_1_cast_fp16 = batch_norm(beta = obj_1_beta_0_to_fp16, epsilon = obj_1_epsilon_0_to_fp16, gamma = obj_1_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_3_cast_fp16)[name = tensor<string, []>("obj_1_cast_fp16")];
             tensor<string, []> query_1_pad_type_0 = const()[name = tensor<string, []>("query_1_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> query_1_strides_0 = const()[name = tensor<string, []>("query_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> query_1_pad_0 = const()[name = tensor<string, []>("query_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> query_1_dilations_0 = const()[name = tensor<string, []>("query_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> query_1_groups_0 = const()[name = tensor<string, []>("query_1_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 1024, 1, 1]> layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25468288)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_1_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = query_1_dilations_0, groups = query_1_groups_0, pad = query_1_pad_0, pad_type = query_1_pad_type_0, strides = query_1_strides_0, weight = layers_0_self_attn_q_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor<string, []>("query_1_cast_fp16")];
+            tensor<fp16, [1024, 1024, 1, 1]> layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26236352)))];
+            tensor<fp16, [1, 1024, 1, 188]> query_1_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_1_dilations_0, groups = query_1_groups_0, pad = query_1_pad_0, pad_type = query_1_pad_type_0, strides = query_1_strides_0, weight = layers_0_self_attn_q_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor<string, []>("query_1_cast_fp16")];
             tensor<string, []> key_1_pad_type_0 = const()[name = tensor<string, []>("key_1_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> key_1_strides_0 = const()[name = tensor<string, []>("key_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> key_1_pad_0 = const()[name = tensor<string, []>("key_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> key_1_dilations_0 = const()[name = tensor<string, []>("key_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> key_1_groups_0 = const()[name = tensor<string, []>("key_1_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 1024, 1, 1]> layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27565504)))];
+            tensor<fp16, [1024, 1024, 1, 1]> layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28333568)))];
             tensor<fp16, [1, 1024, 1, 188]> key_1_cast_fp16 = conv(dilations = key_1_dilations_0, groups = key_1_groups_0, pad = key_1_pad_0, pad_type = key_1_pad_type_0, strides = key_1_strides_0, weight = layers_0_self_attn_k_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor<string, []>("key_1_cast_fp16")];
             tensor<string, []> value_1_pad_type_0 = const()[name = tensor<string, []>("value_1_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> value_1_strides_0 = const()[name = tensor<string, []>("value_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> value_1_pad_0 = const()[name = tensor<string, []>("value_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> value_1_dilations_0 = const()[name = tensor<string, []>("value_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> value_1_groups_0 = const()[name = tensor<string, []>("value_1_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 1024, 1, 1]> layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29662720)))];
-            tensor<fp16, [1, 1024, 1, 188]> value_1_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = value_1_dilations_0, groups = value_1_groups_0, pad = value_1_pad_0, pad_type = value_1_pad_type_0, strides = value_1_strides_0, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor<string, []>("value_1_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_240_to_fp16 = const()[name = tensor<string, []>("op_240_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31759936)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_3_cast_fp16 = add(x = query_1_cast_fp16, y = var_240_to_fp16)[name = tensor<string, []>("query_3_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_243_to_fp16 = const()[name = tensor<string, []>("op_243_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31762048)))];
-            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_1_cast_fp16 = add(x = query_1_cast_fp16, y = var_243_to_fp16)[name = tensor<string, []>("q_with_bias_v_1_cast_fp16")];
+            tensor<fp16, [1024, 1024, 1, 1]> layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30430784)))];
+            tensor<fp16, [1, 1024, 1, 188]> value_1_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_1_dilations_0, groups = value_1_groups_0, pad = value_1_pad_0, pad_type = value_1_pad_type_0, strides = value_1_strides_0, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor<string, []>("value_1_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_238_to_fp16 = const()[name = tensor<string, []>("op_238_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32528000)))];
+            tensor<fp16, [1, 1024, 1, 188]> query_3_cast_fp16 = add(x = query_1_cast_fp16, y = var_238_to_fp16)[name = tensor<string, []>("query_3_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_241_to_fp16 = const()[name = tensor<string, []>("op_241_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32530112)))];
+            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_1_cast_fp16 = add(x = query_1_cast_fp16, y = var_241_to_fp16)[name = tensor<string, []>("q_with_bias_v_1_cast_fp16")];
             tensor<string, []> p_1_pad_type_0 = const()[name = tensor<string, []>("p_1_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> p_1_strides_0 = const()[name = tensor<string, []>("p_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> p_1_pad_0 = const()[name = tensor<string, []>("p_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> p_1_dilations_0 = const()[name = tensor<string, []>("p_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> p_1_groups_0 = const()[name = tensor<string, []>("p_1_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1, 1024, 1, 375]> pos_enc_to_fp16 = const()[name = tensor<string, []>("pos_enc_to_fp16"), val = tensor<fp16, [1, 1024, 1, 375]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31764160)))];
             tensor<fp16, [1024, 1024, 1, 1]> layers_0_self_attn_linear_pos_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_linear_pos_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32532224)))];
-            tensor<fp16, [1, 1024, 1, 375]> p_1_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = p_1_dilations_0, groups = p_1_groups_0, pad = p_1_pad_0, pad_type = p_1_pad_type_0, strides = p_1_strides_0, weight = layers_0_self_attn_linear_pos_weight_to_fp16, x = pos_enc_to_fp16)[name = tensor<string, []>("p_1_cast_fp16")];
-            tensor<int32, [4]> var_254 = const()[name = tensor<string, []>("op_254"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_255_cast_fp16 = reshape(shape = var_254, x = q_with_bias_v_1_cast_fp16)[name = tensor<string, []>("op_255_cast_fp16")];
-            tensor<int32, [4]> var_256 = const()[name = tensor<string, []>("op_256"), val = tensor<int32, [4]>([1, 8, 128, -1])];
-            tensor<fp16, [1, 8, 128, 375]> var_257_cast_fp16 = reshape(shape = var_256, x = p_1_cast_fp16)[name = tensor<string, []>("op_257_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 375]> p_1_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_1_dilations_0, groups = p_1_groups_0, pad = p_1_pad_0, pad_type = p_1_pad_type_0, strides = p_1_strides_0, weight = layers_0_self_attn_linear_pos_weight_to_fp16, x = obj_3_cast_fp16)[name = tensor<string, []>("p_1_cast_fp16")];
+            tensor<int32, [4]> var_252 = const()[name = tensor<string, []>("op_252"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_253_cast_fp16 = reshape(shape = var_252, x = q_with_bias_v_1_cast_fp16)[name = tensor<string, []>("op_253_cast_fp16")];
+            tensor<int32, [4]> var_254 = const()[name = tensor<string, []>("op_254"), val = tensor<int32, [4]>([1, 8, 128, -1])];
+            tensor<fp16, [1, 8, 128, 375]> var_255_cast_fp16 = reshape(shape = var_254, x = p_1_cast_fp16)[name = tensor<string, []>("op_255_cast_fp16")];
             tensor<bool, []> matrix_bd_1_transpose_x_0 = const()[name = tensor<string, []>("matrix_bd_1_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> matrix_bd_1_transpose_y_0 = const()[name = tensor<string, []>("matrix_bd_1_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_1_cast_fp16 = matmul(transpose_x = matrix_bd_1_transpose_x_0, transpose_y = matrix_bd_1_transpose_y_0, x = var_255_cast_fp16, y = var_257_cast_fp16)[name = tensor<string, []>("matrix_bd_1_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_1_cast_fp16 = matmul(transpose_x = matrix_bd_1_transpose_x_0, transpose_y = matrix_bd_1_transpose_y_0, x = var_253_cast_fp16, y = var_255_cast_fp16)[name = tensor<string, []>("matrix_bd_1_cast_fp16")];
             tensor<int32, [8]> matrix_bd_3_pad_0 = const()[name = tensor<string, []>("matrix_bd_3_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
             tensor<string, []> matrix_bd_3_mode_0 = const()[name = tensor<string, []>("matrix_bd_3_mode_0"), val = tensor<string, []>("constant")];
             tensor<fp16, []> const_10_to_fp16 = const()[name = tensor<string, []>("const_10_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
             tensor<fp16, [1, 8, 188, 376]> matrix_bd_3_cast_fp16 = pad(constant_val = const_10_to_fp16, mode = matrix_bd_3_mode_0, pad = matrix_bd_3_pad_0, x = matrix_bd_1_cast_fp16)[name = tensor<string, []>("matrix_bd_3_cast_fp16")];
-            tensor<int32, [4]> var_266 = const()[name = tensor<string, []>("op_266"), val = tensor<int32, [4]>([1, 8, -1, 188])];
-            tensor<fp16, [1, 8, 376, 188]> matrix_bd_5_cast_fp16 = reshape(shape = var_266, x = matrix_bd_3_cast_fp16)[name = tensor<string, []>("matrix_bd_5_cast_fp16")];
-            tensor<int32, [4]> var_270_begin_0 = const()[name = tensor<string, []>("op_270_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
-            tensor<int32, [4]> var_270_end_0 = const()[name = tensor<string, []>("op_270_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
-            tensor<bool, [4]> var_270_end_mask_0 = const()[name = tensor<string, []>("op_270_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
-            tensor<fp16, [1, 8, 375, 188]> var_270_cast_fp16 = slice_by_index(begin = var_270_begin_0, end = var_270_end_0, end_mask = var_270_end_mask_0, x = matrix_bd_5_cast_fp16)[name = tensor<string, []>("op_270_cast_fp16")];
-            tensor<int32, [4]> var_271 = const()[name = tensor<string, []>("op_271"), val = tensor<int32, [4]>([1, 8, 188, 375])];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_7_cast_fp16 = reshape(shape = var_271, x = var_270_cast_fp16)[name = tensor<string, []>("matrix_bd_7_cast_fp16")];
-            tensor<int32, [4]> var_276_begin_0 = const()[name = tensor<string, []>("op_276_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [4]> var_276_end_0 = const()[name = tensor<string, []>("op_276_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
-            tensor<bool, [4]> var_276_end_mask_0 = const()[name = tensor<string, []>("op_276_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
-            tensor<fp16, [1, 8, 188, 188]> var_276_cast_fp16 = slice_by_index(begin = var_276_begin_0, end = var_276_end_0, end_mask = var_276_end_mask_0, x = matrix_bd_7_cast_fp16)[name = tensor<string, []>("op_276_cast_fp16")];
-            tensor<fp16, []> var_277_to_fp16 = const()[name = tensor<string, []>("op_277_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 188, 188]> qk_mask_1_cast_fp16 = mul(x = var_276_cast_fp16, y = var_277_to_fp16)[name = tensor<string, []>("qk_mask_1_cast_fp16")];
-            tensor<int32, [4]> var_281 = const()[name = tensor<string, []>("op_281"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> mh_q_1_cast_fp16 = reshape(shape = var_281, x = query_3_cast_fp16)[name = tensor<string, []>("mh_q_1_cast_fp16")];
-            tensor<fp16, []> var_283_to_fp16 = const()[name = tensor<string, []>("op_283_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 128, 188]> var_284_cast_fp16 = mul(x = mh_q_1_cast_fp16, y = var_283_to_fp16)[name = tensor<string, []>("op_284_cast_fp16")];
-            tensor<int32, [4]> var_287 = const()[name = tensor<string, []>("op_287"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_288_cast_fp16 = reshape(shape = var_287, x = key_1_cast_fp16)[name = tensor<string, []>("op_288_cast_fp16")];
+            tensor<int32, [4]> var_264 = const()[name = tensor<string, []>("op_264"), val = tensor<int32, [4]>([1, 8, -1, 188])];
+            tensor<fp16, [1, 8, 376, 188]> matrix_bd_5_cast_fp16 = reshape(shape = var_264, x = matrix_bd_3_cast_fp16)[name = tensor<string, []>("matrix_bd_5_cast_fp16")];
+            tensor<int32, [4]> var_268_begin_0 = const()[name = tensor<string, []>("op_268_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_268_end_0 = const()[name = tensor<string, []>("op_268_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
+            tensor<bool, [4]> var_268_end_mask_0 = const()[name = tensor<string, []>("op_268_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 375, 188]> var_268_cast_fp16 = slice_by_index(begin = var_268_begin_0, end = var_268_end_0, end_mask = var_268_end_mask_0, x = matrix_bd_5_cast_fp16)[name = tensor<string, []>("op_268_cast_fp16")];
+            tensor<int32, [4]> var_269 = const()[name = tensor<string, []>("op_269"), val = tensor<int32, [4]>([1, 8, 188, 375])];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_7_cast_fp16 = reshape(shape = var_269, x = var_268_cast_fp16)[name = tensor<string, []>("matrix_bd_7_cast_fp16")];
+            tensor<int32, [4]> var_274_begin_0 = const()[name = tensor<string, []>("op_274_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_274_end_0 = const()[name = tensor<string, []>("op_274_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
+            tensor<bool, [4]> var_274_end_mask_0 = const()[name = tensor<string, []>("op_274_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 188, 188]> var_274_cast_fp16 = slice_by_index(begin = var_274_begin_0, end = var_274_end_0, end_mask = var_274_end_mask_0, x = matrix_bd_7_cast_fp16)[name = tensor<string, []>("op_274_cast_fp16")];
+            tensor<fp16, []> var_275_to_fp16 = const()[name = tensor<string, []>("op_275_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 188, 188]> qk_mask_1_cast_fp16 = mul(x = var_274_cast_fp16, y = var_275_to_fp16)[name = tensor<string, []>("qk_mask_1_cast_fp16")];
+            tensor<int32, [4]> var_279 = const()[name = tensor<string, []>("op_279"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> mh_q_1_cast_fp16 = reshape(shape = var_279, x = query_3_cast_fp16)[name = tensor<string, []>("mh_q_1_cast_fp16")];
+            tensor<fp16, []> var_281_to_fp16 = const()[name = tensor<string, []>("op_281_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 128, 188]> var_282_cast_fp16 = mul(x = mh_q_1_cast_fp16, y = var_281_to_fp16)[name = tensor<string, []>("op_282_cast_fp16")];
+            tensor<int32, [4]> var_285 = const()[name = tensor<string, []>("op_285"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_286_cast_fp16 = reshape(shape = var_285, x = key_1_cast_fp16)[name = tensor<string, []>("op_286_cast_fp16")];
             tensor<bool, []> mh_w_1_transpose_x_0 = const()[name = tensor<string, []>("mh_w_1_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> mh_w_1_transpose_y_0 = const()[name = tensor<string, []>("mh_w_1_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 188]> mh_w_1_cast_fp16 = matmul(transpose_x = mh_w_1_transpose_x_0, transpose_y = mh_w_1_transpose_y_0, x = var_284_cast_fp16, y = var_288_cast_fp16)[name = tensor<string, []>("mh_w_1_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> mh_w_1_cast_fp16 = matmul(transpose_x = mh_w_1_transpose_x_0, transpose_y = mh_w_1_transpose_y_0, x = var_282_cast_fp16, y = var_286_cast_fp16)[name = tensor<string, []>("mh_w_1_cast_fp16")];
             tensor<fp16, [1, 8, 188, 188]> mh_w_3_cast_fp16 = add(x = mh_w_1_cast_fp16, y = qk_mask_1_cast_fp16)[name = tensor<string, []>("mh_w_3_cast_fp16")];
-            tensor<fp16, [1, 8, 188, 188]> var_292_cast_fp16 = softmax(axis = var_133, x = mh_w_3_cast_fp16)[name = tensor<string, []>("op_292_cast_fp16")];
-            tensor<int32, [4]> var_293 = const()[name = tensor<string, []>("op_293"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_294_cast_fp16 = reshape(shape = var_293, x = value_1_cast_fp16)[name = tensor<string, []>("op_294_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> var_290_cast_fp16 = softmax(axis = var_131, x = mh_w_3_cast_fp16)[name = tensor<string, []>("op_290_cast_fp16")];
+            tensor<int32, [4]> var_291 = const()[name = tensor<string, []>("op_291"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_292_cast_fp16 = reshape(shape = var_291, x = value_1_cast_fp16)[name = tensor<string, []>("op_292_cast_fp16")];
             tensor<bool, []> attn_1_transpose_x_0 = const()[name = tensor<string, []>("attn_1_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> attn_1_transpose_y_0 = const()[name = tensor<string, []>("attn_1_transpose_y_0"), val = tensor<bool, []>(true)];
-            tensor<fp16, [1, 8, 128, 188]> attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = var_294_cast_fp16, y = var_292_cast_fp16)[name = tensor<string, []>("attn_1_cast_fp16")];
-            tensor<int32, [4]> var_297 = const()[name = tensor<string, []>("op_297"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
-            tensor<fp16, [1, 1024, 1, 188]> input_21_cast_fp16 = reshape(shape = var_297, x = attn_1_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")];
+            tensor<fp16, [1, 8, 128, 188]> attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = var_292_cast_fp16, y = var_290_cast_fp16)[name = tensor<string, []>("attn_1_cast_fp16")];
+            tensor<int32, [4]> var_295 = const()[name = tensor<string, []>("op_295"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
+            tensor<fp16, [1, 1024, 1, 188]> input_23_cast_fp16 = reshape(shape = var_295, x = attn_1_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")];
             tensor<string, []> obj_5_pad_type_0 = const()[name = tensor<string, []>("obj_5_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> obj_5_strides_0 = const()[name = tensor<string, []>("obj_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> obj_5_pad_0 = const()[name = tensor<string, []>("obj_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> obj_5_dilations_0 = const()[name = tensor<string, []>("obj_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> obj_5_groups_0 = const()[name = tensor<string, []>("obj_5_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_0_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34629440)))];
-            tensor<fp16, [1, 1024, 1, 188]> obj_5_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = obj_5_dilations_0, groups = obj_5_groups_0, pad = obj_5_pad_0, pad_type = obj_5_pad_type_0, strides = obj_5_strides_0, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("obj_5_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_5_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_5_dilations_0, groups = obj_5_groups_0, pad = obj_5_pad_0, pad_type = obj_5_pad_type_0, strides = obj_5_strides_0, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = input_23_cast_fp16)[name = tensor<string, []>("obj_5_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = obj_5_cast_fp16)[name = tensor<string, []>("inputs_5_cast_fp16")];
             tensor<int32, [1]> out_5_axes_0 = const()[name = tensor<string, []>("out_5_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_315_to_fp16 = const()[name = tensor<string, []>("op_315_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_315_to_fp16, x = inputs_5_cast_fp16)[name = tensor<string, []>("out_5_cast_fp16")];
-            tensor<fp16, [1024]> input_23_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_23_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36726656)))];
-            tensor<fp16, [1024]> input_23_beta_0_to_fp16 = const()[name = tensor<string, []>("input_23_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36728768)))];
-            tensor<fp16, []> input_23_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_23_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_23_cast_fp16 = batch_norm(beta = input_23_beta_0_to_fp16, epsilon = input_23_epsilon_0_to_fp16, gamma = input_23_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_5_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")];
-            tensor<string, []> input_25_pad_type_0 = const()[name = tensor<string, []>("input_25_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_25_strides_0 = const()[name = tensor<string, []>("input_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_25_pad_0 = const()[name = tensor<string, []>("input_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_25_dilations_0 = const()[name = tensor<string, []>("input_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_25_groups_0 = const()[name = tensor<string, []>("input_25_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_313_to_fp16 = const()[name = tensor<string, []>("op_313_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_313_to_fp16, x = inputs_5_cast_fp16)[name = tensor<string, []>("out_5_cast_fp16")];
+            tensor<fp16, [1024]> input_25_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_25_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36726656)))];
+            tensor<fp16, [1024]> input_25_beta_0_to_fp16 = const()[name = tensor<string, []>("input_25_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36728768)))];
+            tensor<fp16, []> input_25_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_25_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_25_cast_fp16 = batch_norm(beta = input_25_beta_0_to_fp16, epsilon = input_25_epsilon_0_to_fp16, gamma = input_25_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_5_cast_fp16)[name = tensor<string, []>("input_25_cast_fp16")];
+            tensor<string, []> input_27_pad_type_0 = const()[name = tensor<string, []>("input_27_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_27_strides_0 = const()[name = tensor<string, []>("input_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_27_pad_0 = const()[name = tensor<string, []>("input_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_27_dilations_0 = const()[name = tensor<string, []>("input_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_27_groups_0 = const()[name = tensor<string, []>("input_27_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [2048, 1024, 1, 1]> layers_0_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36730880)))];
-            tensor<fp16, [1, 2048, 1, 188]> input_25_cast_fp16 = conv(dilations = input_25_dilations_0, groups = input_25_groups_0, pad = input_25_pad_0, pad_type = input_25_pad_type_0, strides = input_25_strides_0, weight = layers_0_conv_pointwise_conv1_weight_to_fp16, x = input_23_cast_fp16)[name = tensor<string, []>("input_25_cast_fp16")];
-            tensor<int32, []> input_27_split_num_splits_0 = const()[name = tensor<string, []>("input_27_split_num_splits_0"), val = tensor<int32, []>(2)];
-            tensor<int32, []> input_27_split_axis_0 = const()[name = tensor<string, []>("input_27_split_axis_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1, 1024, 1, 188]> input_27_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_27_split_cast_fp16_1 = split(axis = input_27_split_axis_0, num_splits = input_27_split_num_splits_0, x = input_25_cast_fp16)[name = tensor<string, []>("input_27_split_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_27_split_1_sigmoid_cast_fp16 = sigmoid(x = input_27_split_cast_fp16_1)[name = tensor<string, []>("input_27_split_1_sigmoid_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_27_cast_fp16 = mul(x = input_27_split_cast_fp16_0, y = input_27_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")];
-            tensor<string, []> input_29_pad_type_0 = const()[name = tensor<string, []>("input_29_pad_type_0"), val = tensor<string, []>("custom")];
-            tensor<int32, [4]> input_29_pad_0 = const()[name = tensor<string, []>("input_29_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
-            tensor<int32, []> input_29_groups_0 = const()[name = tensor<string, []>("input_29_groups_0"), val = tensor<int32, []>(1024)];
-            tensor<int32, [2]> input_29_strides_0 = const()[name = tensor<string, []>("input_29_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> input_29_dilations_0 = const()[name = tensor<string, []>("input_29_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [1, 2048, 1, 188]> input_27_cast_fp16 = conv(dilations = input_27_dilations_0, groups = input_27_groups_0, pad = input_27_pad_0, pad_type = input_27_pad_type_0, strides = input_27_strides_0, weight = layers_0_conv_pointwise_conv1_weight_to_fp16, x = input_25_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")];
+            tensor<int32, []> input_29_split_num_splits_0 = const()[name = tensor<string, []>("input_29_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> input_29_split_axis_0 = const()[name = tensor<string, []>("input_29_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 1, 188]> input_29_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_29_split_cast_fp16_1 = split(axis = input_29_split_axis_0, num_splits = input_29_split_num_splits_0, x = input_27_cast_fp16)[name = tensor<string, []>("input_29_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_29_split_1_sigmoid_cast_fp16 = sigmoid(x = input_29_split_cast_fp16_1)[name = tensor<string, []>("input_29_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_29_cast_fp16 = mul(x = input_29_split_cast_fp16_0, y = input_29_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_29_cast_fp16")];
+            tensor<string, []> input_31_pad_type_0 = const()[name = tensor<string, []>("input_31_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_31_pad_0 = const()[name = tensor<string, []>("input_31_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
+            tensor<int32, []> input_31_groups_0 = const()[name = tensor<string, []>("input_31_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [2]> input_31_strides_0 = const()[name = tensor<string, []>("input_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> input_31_dilations_0 = const()[name = tensor<string, []>("input_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<fp16, [1024, 1, 1, 9]> const_268_to_fp16 = const()[name = tensor<string, []>("const_268_to_fp16"), val = tensor<fp16, [1024, 1, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40925248)))];
             tensor<fp16, [1024]> const_269_to_fp16 = const()[name = tensor<string, []>("const_269_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40943744)))];
-            tensor<fp16, [1, 1024, 1, 188]> input_31_cast_fp16 = conv(bias = const_269_to_fp16, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = const_268_to_fp16, x = input_27_cast_fp16)[name = tensor<string, []>("input_31_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_33_cast_fp16 = silu(x = input_31_cast_fp16)[name = tensor<string, []>("input_33_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_33_cast_fp16 = conv(bias = const_269_to_fp16, dilations = input_31_dilations_0, groups = input_31_groups_0, pad = input_31_pad_0, pad_type = input_31_pad_type_0, strides = input_31_strides_0, weight = const_268_to_fp16, x = input_29_cast_fp16)[name = tensor<string, []>("input_33_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_35_cast_fp16 = silu(x = input_33_cast_fp16)[name = tensor<string, []>("input_35_cast_fp16")];
             tensor<string, []> x_5_pad_type_0 = const()[name = tensor<string, []>("x_5_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_5_strides_0 = const()[name = tensor<string, []>("x_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_5_pad_0 = const()[name = tensor<string, []>("x_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_5_dilations_0 = const()[name = tensor<string, []>("x_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_5_groups_0 = const()[name = tensor<string, []>("x_5_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_0_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40945856)))];
-            tensor<fp16, [1, 1024, 1, 188]> x_5_cast_fp16 = conv(dilations = x_5_dilations_0, groups = x_5_groups_0, pad = x_5_pad_0, pad_type = x_5_pad_type_0, strides = x_5_strides_0, weight = layers_0_conv_pointwise_conv2_weight_to_fp16, x = input_33_cast_fp16)[name = tensor<string, []>("x_5_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> x_5_cast_fp16 = conv(dilations = x_5_dilations_0, groups = x_5_groups_0, pad = x_5_pad_0, pad_type = x_5_pad_type_0, strides = x_5_strides_0, weight = layers_0_conv_pointwise_conv2_weight_to_fp16, x = input_35_cast_fp16)[name = tensor<string, []>("x_5_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = x_5_cast_fp16)[name = tensor<string, []>("inputs_7_cast_fp16")];
             tensor<int32, [1]> out_7_axes_0 = const()[name = tensor<string, []>("out_7_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_363_to_fp16 = const()[name = tensor<string, []>("op_363_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_363_to_fp16, x = inputs_7_cast_fp16)[name = tensor<string, []>("out_7_cast_fp16")];
-            tensor<fp16, [1024]> input_35_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_35_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43043072)))];
-            tensor<fp16, [1024]> input_35_beta_0_to_fp16 = const()[name = tensor<string, []>("input_35_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43045184)))];
-            tensor<fp16, []> input_35_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_35_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_35_cast_fp16 = batch_norm(beta = input_35_beta_0_to_fp16, epsilon = input_35_epsilon_0_to_fp16, gamma = input_35_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_7_cast_fp16)[name = tensor<string, []>("input_35_cast_fp16")];
-            tensor<string, []> input_37_pad_type_0 = const()[name = tensor<string, []>("input_37_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_37_strides_0 = const()[name = tensor<string, []>("input_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_37_pad_0 = const()[name = tensor<string, []>("input_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_37_dilations_0 = const()[name = tensor<string, []>("input_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_37_groups_0 = const()[name = tensor<string, []>("input_37_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_361_to_fp16 = const()[name = tensor<string, []>("op_361_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_361_to_fp16, x = inputs_7_cast_fp16)[name = tensor<string, []>("out_7_cast_fp16")];
+            tensor<fp16, [1024]> input_37_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_37_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43043072)))];
+            tensor<fp16, [1024]> input_37_beta_0_to_fp16 = const()[name = tensor<string, []>("input_37_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43045184)))];
+            tensor<fp16, []> input_37_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_37_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_37_cast_fp16 = batch_norm(beta = input_37_beta_0_to_fp16, epsilon = input_37_epsilon_0_to_fp16, gamma = input_37_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_7_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")];
+            tensor<string, []> input_39_pad_type_0 = const()[name = tensor<string, []>("input_39_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_39_strides_0 = const()[name = tensor<string, []>("input_39_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_39_pad_0 = const()[name = tensor<string, []>("input_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_39_dilations_0 = const()[name = tensor<string, []>("input_39_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_39_groups_0 = const()[name = tensor<string, []>("input_39_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_0_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_feed_forward2_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43047296)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_37_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_37_dilations_0, groups = input_37_groups_0, pad = input_37_pad_0, pad_type = input_37_pad_type_0, strides = input_37_strides_0, weight = layers_0_feed_forward2_fc1_weight_to_fp16, x = input_35_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_39_cast_fp16 = silu(x = input_37_cast_fp16)[name = tensor<string, []>("input_39_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_39_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_39_dilations_0, groups = input_39_groups_0, pad = input_39_pad_0, pad_type = input_39_pad_type_0, strides = input_39_strides_0, weight = layers_0_feed_forward2_fc1_weight_to_fp16, x = input_37_cast_fp16)[name = tensor<string, []>("input_39_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_41_cast_fp16 = silu(x = input_39_cast_fp16)[name = tensor<string, []>("input_41_cast_fp16")];
             tensor<string, []> x_7_pad_type_0 = const()[name = tensor<string, []>("x_7_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_7_strides_0 = const()[name = tensor<string, []>("x_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_7_pad_0 = const()[name = tensor<string, []>("x_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_7_dilations_0 = const()[name = tensor<string, []>("x_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_7_groups_0 = const()[name = tensor<string, []>("x_7_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_391_weight_0_to_fp16 = const()[name = tensor<string, []>("op_391_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51435968)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_391_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_7_dilations_0, groups = x_7_groups_0, pad = x_7_pad_0, pad_type = x_7_pad_type_0, strides = x_7_strides_0, weight = var_391_weight_0_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("op_391_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = var_391_cast_fp16)[name = tensor<string, []>("inputs_9_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_389_weight_0_to_fp16 = const()[name = tensor<string, []>("op_389_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51435968)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_389_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_7_dilations_0, groups = x_7_groups_0, pad = x_7_pad_0, pad_type = x_7_pad_type_0, strides = x_7_strides_0, weight = var_389_weight_0_to_fp16, x = input_41_cast_fp16)[name = tensor<string, []>("op_389_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = var_389_cast_fp16)[name = tensor<string, []>("inputs_9_cast_fp16")];
             tensor<int32, [1]> out_9_axes_0 = const()[name = tensor<string, []>("out_9_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_401_to_fp16 = const()[name = tensor<string, []>("op_401_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_401_to_fp16, x = inputs_9_cast_fp16)[name = tensor<string, []>("out_9_cast_fp16")];
+            tensor<fp16, []> var_399_to_fp16 = const()[name = tensor<string, []>("op_399_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_399_to_fp16, x = inputs_9_cast_fp16)[name = tensor<string, []>("out_9_cast_fp16")];
             tensor<fp16, [1024]> inputs_11_gamma_0_to_fp16 = const()[name = tensor<string, []>("inputs_11_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59824640)))];
             tensor<fp16, [1024]> inputs_11_beta_0_to_fp16 = const()[name = tensor<string, []>("inputs_11_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59826752)))];
             tensor<fp16, []> inputs_11_epsilon_0_to_fp16 = const()[name = tensor<string, []>("inputs_11_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_11_cast_fp16 = batch_norm(beta = inputs_11_beta_0_to_fp16, epsilon = inputs_11_epsilon_0_to_fp16, gamma = inputs_11_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_9_cast_fp16)[name = tensor<string, []>("inputs_11_cast_fp16")];
-            tensor<int32, []> var_415 = const()[name = tensor<string, []>("op_415"), val = tensor<int32, []>(3)];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_11_cast_fp16 = batch_norm(beta = inputs_11_beta_0_to_fp16, epsilon = inputs_11_epsilon_0_to_fp16, gamma = inputs_11_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_9_cast_fp16)[name = tensor<string, []>("inputs_11_cast_fp16")];
+            tensor<int32, []> var_413 = const()[name = tensor<string, []>("op_413"), val = tensor<int32, []>(3)];
             tensor<int32, [1]> out_11_axes_0 = const()[name = tensor<string, []>("out_11_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_446_to_fp16 = const()[name = tensor<string, []>("op_446_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_446_to_fp16, x = inputs_11_cast_fp16)[name = tensor<string, []>("out_11_cast_fp16")];
-            tensor<fp16, [1024]> input_41_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_41_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59828864)))];
-            tensor<fp16, [1024]> input_41_beta_0_to_fp16 = const()[name = tensor<string, []>("input_41_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59830976)))];
-            tensor<fp16, []> input_41_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_41_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_41_cast_fp16 = batch_norm(beta = input_41_beta_0_to_fp16, epsilon = input_41_epsilon_0_to_fp16, gamma = input_41_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_11_cast_fp16)[name = tensor<string, []>("input_41_cast_fp16")];
-            tensor<string, []> input_43_pad_type_0 = const()[name = tensor<string, []>("input_43_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_43_strides_0 = const()[name = tensor<string, []>("input_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_43_pad_0 = const()[name = tensor<string, []>("input_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_43_dilations_0 = const()[name = tensor<string, []>("input_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_43_groups_0 = const()[name = tensor<string, []>("input_43_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_444_to_fp16 = const()[name = tensor<string, []>("op_444_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_444_to_fp16, x = inputs_11_cast_fp16)[name = tensor<string, []>("out_11_cast_fp16")];
+            tensor<fp16, [1024]> input_43_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_43_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59828864)))];
+            tensor<fp16, [1024]> input_43_beta_0_to_fp16 = const()[name = tensor<string, []>("input_43_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59830976)))];
+            tensor<fp16, []> input_43_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_43_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_43_cast_fp16 = batch_norm(beta = input_43_beta_0_to_fp16, epsilon = input_43_epsilon_0_to_fp16, gamma = input_43_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_11_cast_fp16)[name = tensor<string, []>("input_43_cast_fp16")];
+            tensor<string, []> input_45_pad_type_0 = const()[name = tensor<string, []>("input_45_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_45_strides_0 = const()[name = tensor<string, []>("input_45_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_45_pad_0 = const()[name = tensor<string, []>("input_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_45_dilations_0 = const()[name = tensor<string, []>("input_45_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_45_groups_0 = const()[name = tensor<string, []>("input_45_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_1_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_feed_forward1_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59833088)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_43_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_43_dilations_0, groups = input_43_groups_0, pad = input_43_pad_0, pad_type = input_43_pad_type_0, strides = input_43_strides_0, weight = layers_1_feed_forward1_fc1_weight_to_fp16, x = input_41_cast_fp16)[name = tensor<string, []>("input_43_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_45_cast_fp16 = silu(x = input_43_cast_fp16)[name = tensor<string, []>("input_45_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_45_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_45_dilations_0, groups = input_45_groups_0, pad = input_45_pad_0, pad_type = input_45_pad_type_0, strides = input_45_strides_0, weight = layers_1_feed_forward1_fc1_weight_to_fp16, x = input_43_cast_fp16)[name = tensor<string, []>("input_45_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_47_cast_fp16 = silu(x = input_45_cast_fp16)[name = tensor<string, []>("input_47_cast_fp16")];
             tensor<string, []> x_9_pad_type_0 = const()[name = tensor<string, []>("x_9_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_9_strides_0 = const()[name = tensor<string, []>("x_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_9_pad_0 = const()[name = tensor<string, []>("x_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_9_dilations_0 = const()[name = tensor<string, []>("x_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_9_groups_0 = const()[name = tensor<string, []>("x_9_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_474_weight_0_to_fp16 = const()[name = tensor<string, []>("op_474_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(68221760)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_474_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_9_dilations_0, groups = x_9_groups_0, pad = x_9_pad_0, pad_type = x_9_pad_type_0, strides = x_9_strides_0, weight = var_474_weight_0_to_fp16, x = input_45_cast_fp16)[name = tensor<string, []>("op_474_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = var_474_cast_fp16)[name = tensor<string, []>("inputs_13_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_472_weight_0_to_fp16 = const()[name = tensor<string, []>("op_472_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(68221760)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_472_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_9_dilations_0, groups = x_9_groups_0, pad = x_9_pad_0, pad_type = x_9_pad_type_0, strides = x_9_strides_0, weight = var_472_weight_0_to_fp16, x = input_47_cast_fp16)[name = tensor<string, []>("op_472_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = var_472_cast_fp16)[name = tensor<string, []>("inputs_13_cast_fp16")];
             tensor<int32, [1]> out_13_axes_0 = const()[name = tensor<string, []>("out_13_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_484_to_fp16 = const()[name = tensor<string, []>("op_484_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_484_to_fp16, x = inputs_13_cast_fp16)[name = tensor<string, []>("out_13_cast_fp16")];
+            tensor<fp16, []> var_482_to_fp16 = const()[name = tensor<string, []>("op_482_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_482_to_fp16, x = inputs_13_cast_fp16)[name = tensor<string, []>("out_13_cast_fp16")];
             tensor<fp16, [1024]> obj_7_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_7_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(76610432)))];
             tensor<fp16, [1024]> obj_7_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_7_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(76612544)))];
             tensor<fp16, []> obj_7_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_7_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> obj_7_cast_fp16 = batch_norm(beta = obj_7_beta_0_to_fp16, epsilon = obj_7_epsilon_0_to_fp16, gamma = obj_7_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_13_cast_fp16)[name = tensor<string, []>("obj_7_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_7_cast_fp16 = batch_norm(beta = obj_7_beta_0_to_fp16, epsilon = obj_7_epsilon_0_to_fp16, gamma = obj_7_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_13_cast_fp16)[name = tensor<string, []>("obj_7_cast_fp16")];
             tensor<string, []> query_5_pad_type_0 = const()[name = tensor<string, []>("query_5_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> query_5_strides_0 = const()[name = tensor<string, []>("query_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> query_5_pad_0 = const()[name = tensor<string, []>("query_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> query_5_dilations_0 = const()[name = tensor<string, []>("query_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> query_5_groups_0 = const()[name = tensor<string, []>("query_5_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(76614656)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_5_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = query_5_dilations_0, groups = query_5_groups_0, pad = query_5_pad_0, pad_type = query_5_pad_type_0, strides = query_5_strides_0, weight = layers_1_self_attn_q_proj_weight_to_fp16, x = obj_7_cast_fp16)[name = tensor<string, []>("query_5_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> query_5_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_5_dilations_0, groups = query_5_groups_0, pad = query_5_pad_0, pad_type = query_5_pad_type_0, strides = query_5_strides_0, weight = layers_1_self_attn_q_proj_weight_to_fp16, x = obj_7_cast_fp16)[name = tensor<string, []>("query_5_cast_fp16")];
             tensor<string, []> key_3_pad_type_0 = const()[name = tensor<string, []>("key_3_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> key_3_strides_0 = const()[name = tensor<string, []>("key_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> key_3_pad_0 = const()[name = tensor<string, []>("key_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
@@ -292,173 +294,173 @@ program(1.0)
             tensor<int32, [2]> value_3_dilations_0 = const()[name = tensor<string, []>("value_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> value_3_groups_0 = const()[name = tensor<string, []>("value_3_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_1_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80809088)))];
-            tensor<fp16, [1, 1024, 1, 188]> value_3_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = value_3_dilations_0, groups = value_3_groups_0, pad = value_3_pad_0, pad_type = value_3_pad_type_0, strides = value_3_strides_0, weight = layers_1_self_attn_v_proj_weight_to_fp16, x = obj_7_cast_fp16)[name = tensor<string, []>("value_3_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_522_to_fp16 = const()[name = tensor<string, []>("op_522_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82906304)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_7_cast_fp16 = add(x = query_5_cast_fp16, y = var_522_to_fp16)[name = tensor<string, []>("query_7_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_525_to_fp16 = const()[name = tensor<string, []>("op_525_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82908416)))];
-            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_3_cast_fp16 = add(x = query_5_cast_fp16, y = var_525_to_fp16)[name = tensor<string, []>("q_with_bias_v_3_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> value_3_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_3_dilations_0, groups = value_3_groups_0, pad = value_3_pad_0, pad_type = value_3_pad_type_0, strides = value_3_strides_0, weight = layers_1_self_attn_v_proj_weight_to_fp16, x = obj_7_cast_fp16)[name = tensor<string, []>("value_3_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_520_to_fp16 = const()[name = tensor<string, []>("op_520_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82906304)))];
+            tensor<fp16, [1, 1024, 1, 188]> query_7_cast_fp16 = add(x = query_5_cast_fp16, y = var_520_to_fp16)[name = tensor<string, []>("query_7_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_523_to_fp16 = const()[name = tensor<string, []>("op_523_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82908416)))];
+            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_3_cast_fp16 = add(x = query_5_cast_fp16, y = var_523_to_fp16)[name = tensor<string, []>("q_with_bias_v_3_cast_fp16")];
             tensor<string, []> p_3_pad_type_0 = const()[name = tensor<string, []>("p_3_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> p_3_strides_0 = const()[name = tensor<string, []>("p_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> p_3_pad_0 = const()[name = tensor<string, []>("p_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> p_3_dilations_0 = const()[name = tensor<string, []>("p_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> p_3_groups_0 = const()[name = tensor<string, []>("p_3_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_1_self_attn_linear_pos_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_linear_pos_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82910528)))];
-            tensor<fp16, [1, 1024, 1, 375]> p_3_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = p_3_dilations_0, groups = p_3_groups_0, pad = p_3_pad_0, pad_type = p_3_pad_type_0, strides = p_3_strides_0, weight = layers_1_self_attn_linear_pos_weight_to_fp16, x = pos_enc_to_fp16)[name = tensor<string, []>("p_3_cast_fp16")];
-            tensor<int32, [4]> var_536 = const()[name = tensor<string, []>("op_536"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_537_cast_fp16 = reshape(shape = var_536, x = q_with_bias_v_3_cast_fp16)[name = tensor<string, []>("op_537_cast_fp16")];
-            tensor<int32, [4]> var_538 = const()[name = tensor<string, []>("op_538"), val = tensor<int32, [4]>([1, 8, 128, -1])];
-            tensor<fp16, [1, 8, 128, 375]> var_539_cast_fp16 = reshape(shape = var_538, x = p_3_cast_fp16)[name = tensor<string, []>("op_539_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 375]> p_3_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_3_dilations_0, groups = p_3_groups_0, pad = p_3_pad_0, pad_type = p_3_pad_type_0, strides = p_3_strides_0, weight = layers_1_self_attn_linear_pos_weight_to_fp16, x = obj_3_cast_fp16)[name = tensor<string, []>("p_3_cast_fp16")];
+            tensor<int32, [4]> var_534 = const()[name = tensor<string, []>("op_534"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_535_cast_fp16 = reshape(shape = var_534, x = q_with_bias_v_3_cast_fp16)[name = tensor<string, []>("op_535_cast_fp16")];
+            tensor<int32, [4]> var_536 = const()[name = tensor<string, []>("op_536"), val = tensor<int32, [4]>([1, 8, 128, -1])];
+            tensor<fp16, [1, 8, 128, 375]> var_537_cast_fp16 = reshape(shape = var_536, x = p_3_cast_fp16)[name = tensor<string, []>("op_537_cast_fp16")];
             tensor<bool, []> matrix_bd_9_transpose_x_0 = const()[name = tensor<string, []>("matrix_bd_9_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> matrix_bd_9_transpose_y_0 = const()[name = tensor<string, []>("matrix_bd_9_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_9_cast_fp16 = matmul(transpose_x = matrix_bd_9_transpose_x_0, transpose_y = matrix_bd_9_transpose_y_0, x = var_537_cast_fp16, y = var_539_cast_fp16)[name = tensor<string, []>("matrix_bd_9_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_9_cast_fp16 = matmul(transpose_x = matrix_bd_9_transpose_x_0, transpose_y = matrix_bd_9_transpose_y_0, x = var_535_cast_fp16, y = var_537_cast_fp16)[name = tensor<string, []>("matrix_bd_9_cast_fp16")];
             tensor<int32, [8]> matrix_bd_11_pad_0 = const()[name = tensor<string, []>("matrix_bd_11_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
             tensor<string, []> matrix_bd_11_mode_0 = const()[name = tensor<string, []>("matrix_bd_11_mode_0"), val = tensor<string, []>("constant")];
             tensor<fp16, []> const_21_to_fp16 = const()[name = tensor<string, []>("const_21_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
             tensor<fp16, [1, 8, 188, 376]> matrix_bd_11_cast_fp16 = pad(constant_val = const_21_to_fp16, mode = matrix_bd_11_mode_0, pad = matrix_bd_11_pad_0, x = matrix_bd_9_cast_fp16)[name = tensor<string, []>("matrix_bd_11_cast_fp16")];
-            tensor<int32, [4]> var_548 = const()[name = tensor<string, []>("op_548"), val = tensor<int32, [4]>([1, 8, -1, 188])];
-            tensor<fp16, [1, 8, 376, 188]> matrix_bd_13_cast_fp16 = reshape(shape = var_548, x = matrix_bd_11_cast_fp16)[name = tensor<string, []>("matrix_bd_13_cast_fp16")];
-            tensor<int32, [4]> var_552_begin_0 = const()[name = tensor<string, []>("op_552_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
-            tensor<int32, [4]> var_552_end_0 = const()[name = tensor<string, []>("op_552_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
-            tensor<bool, [4]> var_552_end_mask_0 = const()[name = tensor<string, []>("op_552_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
-            tensor<fp16, [1, 8, 375, 188]> var_552_cast_fp16 = slice_by_index(begin = var_552_begin_0, end = var_552_end_0, end_mask = var_552_end_mask_0, x = matrix_bd_13_cast_fp16)[name = tensor<string, []>("op_552_cast_fp16")];
-            tensor<int32, [4]> var_553 = const()[name = tensor<string, []>("op_553"), val = tensor<int32, [4]>([1, 8, 188, 375])];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_15_cast_fp16 = reshape(shape = var_553, x = var_552_cast_fp16)[name = tensor<string, []>("matrix_bd_15_cast_fp16")];
-            tensor<int32, [4]> var_558_begin_0 = const()[name = tensor<string, []>("op_558_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [4]> var_558_end_0 = const()[name = tensor<string, []>("op_558_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
-            tensor<bool, [4]> var_558_end_mask_0 = const()[name = tensor<string, []>("op_558_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
-            tensor<fp16, [1, 8, 188, 188]> var_558_cast_fp16 = slice_by_index(begin = var_558_begin_0, end = var_558_end_0, end_mask = var_558_end_mask_0, x = matrix_bd_15_cast_fp16)[name = tensor<string, []>("op_558_cast_fp16")];
-            tensor<fp16, []> var_559_to_fp16 = const()[name = tensor<string, []>("op_559_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 188, 188]> qk_mask_3_cast_fp16 = mul(x = var_558_cast_fp16, y = var_559_to_fp16)[name = tensor<string, []>("qk_mask_3_cast_fp16")];
-            tensor<int32, [4]> var_563 = const()[name = tensor<string, []>("op_563"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> mh_q_3_cast_fp16 = reshape(shape = var_563, x = query_7_cast_fp16)[name = tensor<string, []>("mh_q_3_cast_fp16")];
-            tensor<fp16, []> var_565_to_fp16 = const()[name = tensor<string, []>("op_565_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 128, 188]> var_566_cast_fp16 = mul(x = mh_q_3_cast_fp16, y = var_565_to_fp16)[name = tensor<string, []>("op_566_cast_fp16")];
-            tensor<int32, [4]> var_569 = const()[name = tensor<string, []>("op_569"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_570_cast_fp16 = reshape(shape = var_569, x = key_3_cast_fp16)[name = tensor<string, []>("op_570_cast_fp16")];
+            tensor<int32, [4]> var_546 = const()[name = tensor<string, []>("op_546"), val = tensor<int32, [4]>([1, 8, -1, 188])];
+            tensor<fp16, [1, 8, 376, 188]> matrix_bd_13_cast_fp16 = reshape(shape = var_546, x = matrix_bd_11_cast_fp16)[name = tensor<string, []>("matrix_bd_13_cast_fp16")];
+            tensor<int32, [4]> var_550_begin_0 = const()[name = tensor<string, []>("op_550_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_550_end_0 = const()[name = tensor<string, []>("op_550_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
+            tensor<bool, [4]> var_550_end_mask_0 = const()[name = tensor<string, []>("op_550_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 375, 188]> var_550_cast_fp16 = slice_by_index(begin = var_550_begin_0, end = var_550_end_0, end_mask = var_550_end_mask_0, x = matrix_bd_13_cast_fp16)[name = tensor<string, []>("op_550_cast_fp16")];
+            tensor<int32, [4]> var_551 = const()[name = tensor<string, []>("op_551"), val = tensor<int32, [4]>([1, 8, 188, 375])];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_15_cast_fp16 = reshape(shape = var_551, x = var_550_cast_fp16)[name = tensor<string, []>("matrix_bd_15_cast_fp16")];
+            tensor<int32, [4]> var_556_begin_0 = const()[name = tensor<string, []>("op_556_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_556_end_0 = const()[name = tensor<string, []>("op_556_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
+            tensor<bool, [4]> var_556_end_mask_0 = const()[name = tensor<string, []>("op_556_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 188, 188]> var_556_cast_fp16 = slice_by_index(begin = var_556_begin_0, end = var_556_end_0, end_mask = var_556_end_mask_0, x = matrix_bd_15_cast_fp16)[name = tensor<string, []>("op_556_cast_fp16")];
+            tensor<fp16, []> var_557_to_fp16 = const()[name = tensor<string, []>("op_557_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 188, 188]> qk_mask_3_cast_fp16 = mul(x = var_556_cast_fp16, y = var_557_to_fp16)[name = tensor<string, []>("qk_mask_3_cast_fp16")];
+            tensor<int32, [4]> var_561 = const()[name = tensor<string, []>("op_561"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> mh_q_3_cast_fp16 = reshape(shape = var_561, x = query_7_cast_fp16)[name = tensor<string, []>("mh_q_3_cast_fp16")];
+            tensor<fp16, []> var_563_to_fp16 = const()[name = tensor<string, []>("op_563_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 128, 188]> var_564_cast_fp16 = mul(x = mh_q_3_cast_fp16, y = var_563_to_fp16)[name = tensor<string, []>("op_564_cast_fp16")];
+            tensor<int32, [4]> var_567 = const()[name = tensor<string, []>("op_567"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_568_cast_fp16 = reshape(shape = var_567, x = key_3_cast_fp16)[name = tensor<string, []>("op_568_cast_fp16")];
             tensor<bool, []> mh_w_5_transpose_x_0 = const()[name = tensor<string, []>("mh_w_5_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> mh_w_5_transpose_y_0 = const()[name = tensor<string, []>("mh_w_5_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 188]> mh_w_5_cast_fp16 = matmul(transpose_x = mh_w_5_transpose_x_0, transpose_y = mh_w_5_transpose_y_0, x = var_566_cast_fp16, y = var_570_cast_fp16)[name = tensor<string, []>("mh_w_5_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> mh_w_5_cast_fp16 = matmul(transpose_x = mh_w_5_transpose_x_0, transpose_y = mh_w_5_transpose_y_0, x = var_564_cast_fp16, y = var_568_cast_fp16)[name = tensor<string, []>("mh_w_5_cast_fp16")];
             tensor<fp16, [1, 8, 188, 188]> mh_w_7_cast_fp16 = add(x = mh_w_5_cast_fp16, y = qk_mask_3_cast_fp16)[name = tensor<string, []>("mh_w_7_cast_fp16")];
-            tensor<fp16, [1, 8, 188, 188]> var_574_cast_fp16 = softmax(axis = var_415, x = mh_w_7_cast_fp16)[name = tensor<string, []>("op_574_cast_fp16")];
-            tensor<int32, [4]> var_575 = const()[name = tensor<string, []>("op_575"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_576_cast_fp16 = reshape(shape = var_575, x = value_3_cast_fp16)[name = tensor<string, []>("op_576_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> var_572_cast_fp16 = softmax(axis = var_413, x = mh_w_7_cast_fp16)[name = tensor<string, []>("op_572_cast_fp16")];
+            tensor<int32, [4]> var_573 = const()[name = tensor<string, []>("op_573"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_574_cast_fp16 = reshape(shape = var_573, x = value_3_cast_fp16)[name = tensor<string, []>("op_574_cast_fp16")];
             tensor<bool, []> attn_3_transpose_x_0 = const()[name = tensor<string, []>("attn_3_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> attn_3_transpose_y_0 = const()[name = tensor<string, []>("attn_3_transpose_y_0"), val = tensor<bool, []>(true)];
-            tensor<fp16, [1, 8, 128, 188]> attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = var_576_cast_fp16, y = var_574_cast_fp16)[name = tensor<string, []>("attn_3_cast_fp16")];
-            tensor<int32, [4]> var_579 = const()[name = tensor<string, []>("op_579"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
-            tensor<fp16, [1, 1024, 1, 188]> input_47_cast_fp16 = reshape(shape = var_579, x = attn_3_cast_fp16)[name = tensor<string, []>("input_47_cast_fp16")];
+            tensor<fp16, [1, 8, 128, 188]> attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = var_574_cast_fp16, y = var_572_cast_fp16)[name = tensor<string, []>("attn_3_cast_fp16")];
+            tensor<int32, [4]> var_577 = const()[name = tensor<string, []>("op_577"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
+            tensor<fp16, [1, 1024, 1, 188]> input_49_cast_fp16 = reshape(shape = var_577, x = attn_3_cast_fp16)[name = tensor<string, []>("input_49_cast_fp16")];
             tensor<string, []> obj_9_pad_type_0 = const()[name = tensor<string, []>("obj_9_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> obj_9_strides_0 = const()[name = tensor<string, []>("obj_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> obj_9_pad_0 = const()[name = tensor<string, []>("obj_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> obj_9_dilations_0 = const()[name = tensor<string, []>("obj_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> obj_9_groups_0 = const()[name = tensor<string, []>("obj_9_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_1_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85007744)))];
-            tensor<fp16, [1, 1024, 1, 188]> obj_9_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = obj_9_dilations_0, groups = obj_9_groups_0, pad = obj_9_pad_0, pad_type = obj_9_pad_type_0, strides = obj_9_strides_0, weight = layers_1_self_attn_o_proj_weight_to_fp16, x = input_47_cast_fp16)[name = tensor<string, []>("obj_9_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_9_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_9_dilations_0, groups = obj_9_groups_0, pad = obj_9_pad_0, pad_type = obj_9_pad_type_0, strides = obj_9_strides_0, weight = layers_1_self_attn_o_proj_weight_to_fp16, x = input_49_cast_fp16)[name = tensor<string, []>("obj_9_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = obj_9_cast_fp16)[name = tensor<string, []>("inputs_15_cast_fp16")];
             tensor<int32, [1]> out_15_axes_0 = const()[name = tensor<string, []>("out_15_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_597_to_fp16 = const()[name = tensor<string, []>("op_597_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_597_to_fp16, x = inputs_15_cast_fp16)[name = tensor<string, []>("out_15_cast_fp16")];
-            tensor<fp16, [1024]> input_49_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_49_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87104960)))];
-            tensor<fp16, [1024]> input_49_beta_0_to_fp16 = const()[name = tensor<string, []>("input_49_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87107072)))];
-            tensor<fp16, []> input_49_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_49_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_49_cast_fp16 = batch_norm(beta = input_49_beta_0_to_fp16, epsilon = input_49_epsilon_0_to_fp16, gamma = input_49_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_15_cast_fp16)[name = tensor<string, []>("input_49_cast_fp16")];
-            tensor<string, []> input_51_pad_type_0 = const()[name = tensor<string, []>("input_51_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_51_strides_0 = const()[name = tensor<string, []>("input_51_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_51_pad_0 = const()[name = tensor<string, []>("input_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_51_dilations_0 = const()[name = tensor<string, []>("input_51_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_51_groups_0 = const()[name = tensor<string, []>("input_51_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_595_to_fp16 = const()[name = tensor<string, []>("op_595_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_595_to_fp16, x = inputs_15_cast_fp16)[name = tensor<string, []>("out_15_cast_fp16")];
+            tensor<fp16, [1024]> input_51_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_51_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87104960)))];
+            tensor<fp16, [1024]> input_51_beta_0_to_fp16 = const()[name = tensor<string, []>("input_51_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87107072)))];
+            tensor<fp16, []> input_51_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_51_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_51_cast_fp16 = batch_norm(beta = input_51_beta_0_to_fp16, epsilon = input_51_epsilon_0_to_fp16, gamma = input_51_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_15_cast_fp16)[name = tensor<string, []>("input_51_cast_fp16")];
+            tensor<string, []> input_53_pad_type_0 = const()[name = tensor<string, []>("input_53_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_53_strides_0 = const()[name = tensor<string, []>("input_53_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_53_pad_0 = const()[name = tensor<string, []>("input_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_53_dilations_0 = const()[name = tensor<string, []>("input_53_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_53_groups_0 = const()[name = tensor<string, []>("input_53_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [2048, 1024, 1, 1]> layers_1_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87109184)))];
-            tensor<fp16, [1, 2048, 1, 188]> input_51_cast_fp16 = conv(dilations = input_51_dilations_0, groups = input_51_groups_0, pad = input_51_pad_0, pad_type = input_51_pad_type_0, strides = input_51_strides_0, weight = layers_1_conv_pointwise_conv1_weight_to_fp16, x = input_49_cast_fp16)[name = tensor<string, []>("input_51_cast_fp16")];
-            tensor<int32, []> input_53_split_num_splits_0 = const()[name = tensor<string, []>("input_53_split_num_splits_0"), val = tensor<int32, []>(2)];
-            tensor<int32, []> input_53_split_axis_0 = const()[name = tensor<string, []>("input_53_split_axis_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1, 1024, 1, 188]> input_53_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_53_split_cast_fp16_1 = split(axis = input_53_split_axis_0, num_splits = input_53_split_num_splits_0, x = input_51_cast_fp16)[name = tensor<string, []>("input_53_split_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_53_split_1_sigmoid_cast_fp16 = sigmoid(x = input_53_split_cast_fp16_1)[name = tensor<string, []>("input_53_split_1_sigmoid_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_53_cast_fp16 = mul(x = input_53_split_cast_fp16_0, y = input_53_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_53_cast_fp16")];
-            tensor<string, []> input_55_pad_type_0 = const()[name = tensor<string, []>("input_55_pad_type_0"), val = tensor<string, []>("custom")];
-            tensor<int32, [4]> input_55_pad_0 = const()[name = tensor<string, []>("input_55_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
-            tensor<int32, []> input_55_groups_0 = const()[name = tensor<string, []>("input_55_groups_0"), val = tensor<int32, []>(1024)];
-            tensor<int32, [2]> input_55_strides_0 = const()[name = tensor<string, []>("input_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> input_55_dilations_0 = const()[name = tensor<string, []>("input_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [1, 2048, 1, 188]> input_53_cast_fp16 = conv(dilations = input_53_dilations_0, groups = input_53_groups_0, pad = input_53_pad_0, pad_type = input_53_pad_type_0, strides = input_53_strides_0, weight = layers_1_conv_pointwise_conv1_weight_to_fp16, x = input_51_cast_fp16)[name = tensor<string, []>("input_53_cast_fp16")];
+            tensor<int32, []> input_55_split_num_splits_0 = const()[name = tensor<string, []>("input_55_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> input_55_split_axis_0 = const()[name = tensor<string, []>("input_55_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 1, 188]> input_55_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_55_split_cast_fp16_1 = split(axis = input_55_split_axis_0, num_splits = input_55_split_num_splits_0, x = input_53_cast_fp16)[name = tensor<string, []>("input_55_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_55_split_1_sigmoid_cast_fp16 = sigmoid(x = input_55_split_cast_fp16_1)[name = tensor<string, []>("input_55_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_55_cast_fp16 = mul(x = input_55_split_cast_fp16_0, y = input_55_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_55_cast_fp16")];
+            tensor<string, []> input_57_pad_type_0 = const()[name = tensor<string, []>("input_57_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_57_pad_0 = const()[name = tensor<string, []>("input_57_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
+            tensor<int32, []> input_57_groups_0 = const()[name = tensor<string, []>("input_57_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [2]> input_57_strides_0 = const()[name = tensor<string, []>("input_57_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> input_57_dilations_0 = const()[name = tensor<string, []>("input_57_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<fp16, [1024, 1, 1, 9]> const_270_to_fp16 = const()[name = tensor<string, []>("const_270_to_fp16"), val = tensor<fp16, [1024, 1, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(91303552)))];
             tensor<fp16, [1024]> const_271_to_fp16 = const()[name = tensor<string, []>("const_271_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(91322048)))];
-            tensor<fp16, [1, 1024, 1, 188]> input_57_cast_fp16 = conv(bias = const_271_to_fp16, dilations = input_55_dilations_0, groups = input_55_groups_0, pad = input_55_pad_0, pad_type = input_55_pad_type_0, strides = input_55_strides_0, weight = const_270_to_fp16, x = input_53_cast_fp16)[name = tensor<string, []>("input_57_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_59_cast_fp16 = silu(x = input_57_cast_fp16)[name = tensor<string, []>("input_59_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_59_cast_fp16 = conv(bias = const_271_to_fp16, dilations = input_57_dilations_0, groups = input_57_groups_0, pad = input_57_pad_0, pad_type = input_57_pad_type_0, strides = input_57_strides_0, weight = const_270_to_fp16, x = input_55_cast_fp16)[name = tensor<string, []>("input_59_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_61_cast_fp16 = silu(x = input_59_cast_fp16)[name = tensor<string, []>("input_61_cast_fp16")];
             tensor<string, []> x_11_pad_type_0 = const()[name = tensor<string, []>("x_11_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_11_strides_0 = const()[name = tensor<string, []>("x_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_11_pad_0 = const()[name = tensor<string, []>("x_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_11_dilations_0 = const()[name = tensor<string, []>("x_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_11_groups_0 = const()[name = tensor<string, []>("x_11_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_1_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(91324160)))];
-            tensor<fp16, [1, 1024, 1, 188]> x_11_cast_fp16 = conv(dilations = x_11_dilations_0, groups = x_11_groups_0, pad = x_11_pad_0, pad_type = x_11_pad_type_0, strides = x_11_strides_0, weight = layers_1_conv_pointwise_conv2_weight_to_fp16, x = input_59_cast_fp16)[name = tensor<string, []>("x_11_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> x_11_cast_fp16 = conv(dilations = x_11_dilations_0, groups = x_11_groups_0, pad = x_11_pad_0, pad_type = x_11_pad_type_0, strides = x_11_strides_0, weight = layers_1_conv_pointwise_conv2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor<string, []>("x_11_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = x_11_cast_fp16)[name = tensor<string, []>("inputs_17_cast_fp16")];
             tensor<int32, [1]> out_17_axes_0 = const()[name = tensor<string, []>("out_17_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_645_to_fp16 = const()[name = tensor<string, []>("op_645_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_645_to_fp16, x = inputs_17_cast_fp16)[name = tensor<string, []>("out_17_cast_fp16")];
-            tensor<fp16, [1024]> input_61_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_61_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93421376)))];
-            tensor<fp16, [1024]> input_61_beta_0_to_fp16 = const()[name = tensor<string, []>("input_61_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93423488)))];
-            tensor<fp16, []> input_61_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_61_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_61_cast_fp16 = batch_norm(beta = input_61_beta_0_to_fp16, epsilon = input_61_epsilon_0_to_fp16, gamma = input_61_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_17_cast_fp16)[name = tensor<string, []>("input_61_cast_fp16")];
-            tensor<string, []> input_63_pad_type_0 = const()[name = tensor<string, []>("input_63_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_63_strides_0 = const()[name = tensor<string, []>("input_63_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_63_pad_0 = const()[name = tensor<string, []>("input_63_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_63_dilations_0 = const()[name = tensor<string, []>("input_63_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_63_groups_0 = const()[name = tensor<string, []>("input_63_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_643_to_fp16 = const()[name = tensor<string, []>("op_643_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_643_to_fp16, x = inputs_17_cast_fp16)[name = tensor<string, []>("out_17_cast_fp16")];
+            tensor<fp16, [1024]> input_63_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_63_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93421376)))];
+            tensor<fp16, [1024]> input_63_beta_0_to_fp16 = const()[name = tensor<string, []>("input_63_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93423488)))];
+            tensor<fp16, []> input_63_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_63_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_63_cast_fp16 = batch_norm(beta = input_63_beta_0_to_fp16, epsilon = input_63_epsilon_0_to_fp16, gamma = input_63_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_17_cast_fp16)[name = tensor<string, []>("input_63_cast_fp16")];
+            tensor<string, []> input_65_pad_type_0 = const()[name = tensor<string, []>("input_65_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_65_strides_0 = const()[name = tensor<string, []>("input_65_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_65_pad_0 = const()[name = tensor<string, []>("input_65_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_65_dilations_0 = const()[name = tensor<string, []>("input_65_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_65_groups_0 = const()[name = tensor<string, []>("input_65_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_1_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_feed_forward2_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93425600)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_63_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_63_dilations_0, groups = input_63_groups_0, pad = input_63_pad_0, pad_type = input_63_pad_type_0, strides = input_63_strides_0, weight = layers_1_feed_forward2_fc1_weight_to_fp16, x = input_61_cast_fp16)[name = tensor<string, []>("input_63_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_65_cast_fp16 = silu(x = input_63_cast_fp16)[name = tensor<string, []>("input_65_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_65_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_65_dilations_0, groups = input_65_groups_0, pad = input_65_pad_0, pad_type = input_65_pad_type_0, strides = input_65_strides_0, weight = layers_1_feed_forward2_fc1_weight_to_fp16, x = input_63_cast_fp16)[name = tensor<string, []>("input_65_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_67_cast_fp16 = silu(x = input_65_cast_fp16)[name = tensor<string, []>("input_67_cast_fp16")];
             tensor<string, []> x_13_pad_type_0 = const()[name = tensor<string, []>("x_13_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_13_strides_0 = const()[name = tensor<string, []>("x_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_13_pad_0 = const()[name = tensor<string, []>("x_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_13_dilations_0 = const()[name = tensor<string, []>("x_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_13_groups_0 = const()[name = tensor<string, []>("x_13_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_673_weight_0_to_fp16 = const()[name = tensor<string, []>("op_673_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101814272)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_673_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_13_dilations_0, groups = x_13_groups_0, pad = x_13_pad_0, pad_type = x_13_pad_type_0, strides = x_13_strides_0, weight = var_673_weight_0_to_fp16, x = input_65_cast_fp16)[name = tensor<string, []>("op_673_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = var_673_cast_fp16)[name = tensor<string, []>("inputs_19_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_671_weight_0_to_fp16 = const()[name = tensor<string, []>("op_671_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101814272)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_671_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_13_dilations_0, groups = x_13_groups_0, pad = x_13_pad_0, pad_type = x_13_pad_type_0, strides = x_13_strides_0, weight = var_671_weight_0_to_fp16, x = input_67_cast_fp16)[name = tensor<string, []>("op_671_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = var_671_cast_fp16)[name = tensor<string, []>("inputs_19_cast_fp16")];
             tensor<int32, [1]> out_19_axes_0 = const()[name = tensor<string, []>("out_19_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_683_to_fp16 = const()[name = tensor<string, []>("op_683_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_683_to_fp16, x = inputs_19_cast_fp16)[name = tensor<string, []>("out_19_cast_fp16")];
+            tensor<fp16, []> var_681_to_fp16 = const()[name = tensor<string, []>("op_681_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_681_to_fp16, x = inputs_19_cast_fp16)[name = tensor<string, []>("out_19_cast_fp16")];
             tensor<fp16, [1024]> inputs_21_gamma_0_to_fp16 = const()[name = tensor<string, []>("inputs_21_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110202944)))];
             tensor<fp16, [1024]> inputs_21_beta_0_to_fp16 = const()[name = tensor<string, []>("inputs_21_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110205056)))];
             tensor<fp16, []> inputs_21_epsilon_0_to_fp16 = const()[name = tensor<string, []>("inputs_21_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_21_cast_fp16 = batch_norm(beta = inputs_21_beta_0_to_fp16, epsilon = inputs_21_epsilon_0_to_fp16, gamma = inputs_21_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_19_cast_fp16)[name = tensor<string, []>("inputs_21_cast_fp16")];
-            tensor<int32, []> var_697 = const()[name = tensor<string, []>("op_697"), val = tensor<int32, []>(3)];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_21_cast_fp16 = batch_norm(beta = inputs_21_beta_0_to_fp16, epsilon = inputs_21_epsilon_0_to_fp16, gamma = inputs_21_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_19_cast_fp16)[name = tensor<string, []>("inputs_21_cast_fp16")];
+            tensor<int32, []> var_695 = const()[name = tensor<string, []>("op_695"), val = tensor<int32, []>(3)];
             tensor<int32, [1]> out_21_axes_0 = const()[name = tensor<string, []>("out_21_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_728_to_fp16 = const()[name = tensor<string, []>("op_728_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_728_to_fp16, x = inputs_21_cast_fp16)[name = tensor<string, []>("out_21_cast_fp16")];
-            tensor<fp16, [1024]> input_67_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_67_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110207168)))];
-            tensor<fp16, [1024]> input_67_beta_0_to_fp16 = const()[name = tensor<string, []>("input_67_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110209280)))];
-            tensor<fp16, []> input_67_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_67_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_67_cast_fp16 = batch_norm(beta = input_67_beta_0_to_fp16, epsilon = input_67_epsilon_0_to_fp16, gamma = input_67_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_21_cast_fp16)[name = tensor<string, []>("input_67_cast_fp16")];
-            tensor<string, []> input_69_pad_type_0 = const()[name = tensor<string, []>("input_69_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_69_strides_0 = const()[name = tensor<string, []>("input_69_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_69_pad_0 = const()[name = tensor<string, []>("input_69_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_69_dilations_0 = const()[name = tensor<string, []>("input_69_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_69_groups_0 = const()[name = tensor<string, []>("input_69_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_726_to_fp16 = const()[name = tensor<string, []>("op_726_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_726_to_fp16, x = inputs_21_cast_fp16)[name = tensor<string, []>("out_21_cast_fp16")];
+            tensor<fp16, [1024]> input_69_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_69_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110207168)))];
+            tensor<fp16, [1024]> input_69_beta_0_to_fp16 = const()[name = tensor<string, []>("input_69_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110209280)))];
+            tensor<fp16, []> input_69_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_69_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_69_cast_fp16 = batch_norm(beta = input_69_beta_0_to_fp16, epsilon = input_69_epsilon_0_to_fp16, gamma = input_69_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_21_cast_fp16)[name = tensor<string, []>("input_69_cast_fp16")];
+            tensor<string, []> input_71_pad_type_0 = const()[name = tensor<string, []>("input_71_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_71_strides_0 = const()[name = tensor<string, []>("input_71_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_71_pad_0 = const()[name = tensor<string, []>("input_71_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_71_dilations_0 = const()[name = tensor<string, []>("input_71_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_71_groups_0 = const()[name = tensor<string, []>("input_71_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_2_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_feed_forward1_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110211392)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_69_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = layers_2_feed_forward1_fc1_weight_to_fp16, x = input_67_cast_fp16)[name = tensor<string, []>("input_69_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_71_cast_fp16 = silu(x = input_69_cast_fp16)[name = tensor<string, []>("input_71_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_71_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_71_dilations_0, groups = input_71_groups_0, pad = input_71_pad_0, pad_type = input_71_pad_type_0, strides = input_71_strides_0, weight = layers_2_feed_forward1_fc1_weight_to_fp16, x = input_69_cast_fp16)[name = tensor<string, []>("input_71_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_73_cast_fp16 = silu(x = input_71_cast_fp16)[name = tensor<string, []>("input_73_cast_fp16")];
             tensor<string, []> x_15_pad_type_0 = const()[name = tensor<string, []>("x_15_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_15_strides_0 = const()[name = tensor<string, []>("x_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_15_pad_0 = const()[name = tensor<string, []>("x_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_15_dilations_0 = const()[name = tensor<string, []>("x_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_15_groups_0 = const()[name = tensor<string, []>("x_15_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_756_weight_0_to_fp16 = const()[name = tensor<string, []>("op_756_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118600064)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_756_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_15_dilations_0, groups = x_15_groups_0, pad = x_15_pad_0, pad_type = x_15_pad_type_0, strides = x_15_strides_0, weight = var_756_weight_0_to_fp16, x = input_71_cast_fp16)[name = tensor<string, []>("op_756_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = var_756_cast_fp16)[name = tensor<string, []>("inputs_23_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_754_weight_0_to_fp16 = const()[name = tensor<string, []>("op_754_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118600064)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_754_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_15_dilations_0, groups = x_15_groups_0, pad = x_15_pad_0, pad_type = x_15_pad_type_0, strides = x_15_strides_0, weight = var_754_weight_0_to_fp16, x = input_73_cast_fp16)[name = tensor<string, []>("op_754_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = var_754_cast_fp16)[name = tensor<string, []>("inputs_23_cast_fp16")];
             tensor<int32, [1]> out_23_axes_0 = const()[name = tensor<string, []>("out_23_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_766_to_fp16 = const()[name = tensor<string, []>("op_766_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_766_to_fp16, x = inputs_23_cast_fp16)[name = tensor<string, []>("out_23_cast_fp16")];
+            tensor<fp16, []> var_764_to_fp16 = const()[name = tensor<string, []>("op_764_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_764_to_fp16, x = inputs_23_cast_fp16)[name = tensor<string, []>("out_23_cast_fp16")];
             tensor<fp16, [1024]> obj_11_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_11_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126988736)))];
             tensor<fp16, [1024]> obj_11_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_11_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126990848)))];
             tensor<fp16, []> obj_11_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_11_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> obj_11_cast_fp16 = batch_norm(beta = obj_11_beta_0_to_fp16, epsilon = obj_11_epsilon_0_to_fp16, gamma = obj_11_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_23_cast_fp16)[name = tensor<string, []>("obj_11_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_11_cast_fp16 = batch_norm(beta = obj_11_beta_0_to_fp16, epsilon = obj_11_epsilon_0_to_fp16, gamma = obj_11_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_23_cast_fp16)[name = tensor<string, []>("obj_11_cast_fp16")];
             tensor<string, []> query_9_pad_type_0 = const()[name = tensor<string, []>("query_9_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> query_9_strides_0 = const()[name = tensor<string, []>("query_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> query_9_pad_0 = const()[name = tensor<string, []>("query_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> query_9_dilations_0 = const()[name = tensor<string, []>("query_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> query_9_groups_0 = const()[name = tensor<string, []>("query_9_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_2_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126992960)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_9_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = query_9_dilations_0, groups = query_9_groups_0, pad = query_9_pad_0, pad_type = query_9_pad_type_0, strides = query_9_strides_0, weight = layers_2_self_attn_q_proj_weight_to_fp16, x = obj_11_cast_fp16)[name = tensor<string, []>("query_9_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> query_9_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_9_dilations_0, groups = query_9_groups_0, pad = query_9_pad_0, pad_type = query_9_pad_type_0, strides = query_9_strides_0, weight = layers_2_self_attn_q_proj_weight_to_fp16, x = obj_11_cast_fp16)[name = tensor<string, []>("query_9_cast_fp16")];
             tensor<string, []> key_5_pad_type_0 = const()[name = tensor<string, []>("key_5_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> key_5_strides_0 = const()[name = tensor<string, []>("key_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> key_5_pad_0 = const()[name = tensor<string, []>("key_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
@@ -472,173 +474,173 @@ program(1.0)
             tensor<int32, [2]> value_5_dilations_0 = const()[name = tensor<string, []>("value_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> value_5_groups_0 = const()[name = tensor<string, []>("value_5_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_2_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131187392)))];
-            tensor<fp16, [1, 1024, 1, 188]> value_5_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = value_5_dilations_0, groups = value_5_groups_0, pad = value_5_pad_0, pad_type = value_5_pad_type_0, strides = value_5_strides_0, weight = layers_2_self_attn_v_proj_weight_to_fp16, x = obj_11_cast_fp16)[name = tensor<string, []>("value_5_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_804_to_fp16 = const()[name = tensor<string, []>("op_804_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133284608)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_11_cast_fp16 = add(x = query_9_cast_fp16, y = var_804_to_fp16)[name = tensor<string, []>("query_11_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_807_to_fp16 = const()[name = tensor<string, []>("op_807_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133286720)))];
-            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_5_cast_fp16 = add(x = query_9_cast_fp16, y = var_807_to_fp16)[name = tensor<string, []>("q_with_bias_v_5_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> value_5_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_5_dilations_0, groups = value_5_groups_0, pad = value_5_pad_0, pad_type = value_5_pad_type_0, strides = value_5_strides_0, weight = layers_2_self_attn_v_proj_weight_to_fp16, x = obj_11_cast_fp16)[name = tensor<string, []>("value_5_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_802_to_fp16 = const()[name = tensor<string, []>("op_802_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133284608)))];
+            tensor<fp16, [1, 1024, 1, 188]> query_11_cast_fp16 = add(x = query_9_cast_fp16, y = var_802_to_fp16)[name = tensor<string, []>("query_11_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_805_to_fp16 = const()[name = tensor<string, []>("op_805_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133286720)))];
+            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_5_cast_fp16 = add(x = query_9_cast_fp16, y = var_805_to_fp16)[name = tensor<string, []>("q_with_bias_v_5_cast_fp16")];
             tensor<string, []> p_5_pad_type_0 = const()[name = tensor<string, []>("p_5_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> p_5_strides_0 = const()[name = tensor<string, []>("p_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> p_5_pad_0 = const()[name = tensor<string, []>("p_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> p_5_dilations_0 = const()[name = tensor<string, []>("p_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> p_5_groups_0 = const()[name = tensor<string, []>("p_5_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_2_self_attn_linear_pos_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_linear_pos_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133288832)))];
-            tensor<fp16, [1, 1024, 1, 375]> p_5_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = p_5_dilations_0, groups = p_5_groups_0, pad = p_5_pad_0, pad_type = p_5_pad_type_0, strides = p_5_strides_0, weight = layers_2_self_attn_linear_pos_weight_to_fp16, x = pos_enc_to_fp16)[name = tensor<string, []>("p_5_cast_fp16")];
-            tensor<int32, [4]> var_818 = const()[name = tensor<string, []>("op_818"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_819_cast_fp16 = reshape(shape = var_818, x = q_with_bias_v_5_cast_fp16)[name = tensor<string, []>("op_819_cast_fp16")];
-            tensor<int32, [4]> var_820 = const()[name = tensor<string, []>("op_820"), val = tensor<int32, [4]>([1, 8, 128, -1])];
-            tensor<fp16, [1, 8, 128, 375]> var_821_cast_fp16 = reshape(shape = var_820, x = p_5_cast_fp16)[name = tensor<string, []>("op_821_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 375]> p_5_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_5_dilations_0, groups = p_5_groups_0, pad = p_5_pad_0, pad_type = p_5_pad_type_0, strides = p_5_strides_0, weight = layers_2_self_attn_linear_pos_weight_to_fp16, x = obj_3_cast_fp16)[name = tensor<string, []>("p_5_cast_fp16")];
+            tensor<int32, [4]> var_816 = const()[name = tensor<string, []>("op_816"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_817_cast_fp16 = reshape(shape = var_816, x = q_with_bias_v_5_cast_fp16)[name = tensor<string, []>("op_817_cast_fp16")];
+            tensor<int32, [4]> var_818 = const()[name = tensor<string, []>("op_818"), val = tensor<int32, [4]>([1, 8, 128, -1])];
+            tensor<fp16, [1, 8, 128, 375]> var_819_cast_fp16 = reshape(shape = var_818, x = p_5_cast_fp16)[name = tensor<string, []>("op_819_cast_fp16")];
             tensor<bool, []> matrix_bd_17_transpose_x_0 = const()[name = tensor<string, []>("matrix_bd_17_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> matrix_bd_17_transpose_y_0 = const()[name = tensor<string, []>("matrix_bd_17_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_17_cast_fp16 = matmul(transpose_x = matrix_bd_17_transpose_x_0, transpose_y = matrix_bd_17_transpose_y_0, x = var_819_cast_fp16, y = var_821_cast_fp16)[name = tensor<string, []>("matrix_bd_17_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_17_cast_fp16 = matmul(transpose_x = matrix_bd_17_transpose_x_0, transpose_y = matrix_bd_17_transpose_y_0, x = var_817_cast_fp16, y = var_819_cast_fp16)[name = tensor<string, []>("matrix_bd_17_cast_fp16")];
             tensor<int32, [8]> matrix_bd_19_pad_0 = const()[name = tensor<string, []>("matrix_bd_19_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
             tensor<string, []> matrix_bd_19_mode_0 = const()[name = tensor<string, []>("matrix_bd_19_mode_0"), val = tensor<string, []>("constant")];
             tensor<fp16, []> const_32_to_fp16 = const()[name = tensor<string, []>("const_32_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
             tensor<fp16, [1, 8, 188, 376]> matrix_bd_19_cast_fp16 = pad(constant_val = const_32_to_fp16, mode = matrix_bd_19_mode_0, pad = matrix_bd_19_pad_0, x = matrix_bd_17_cast_fp16)[name = tensor<string, []>("matrix_bd_19_cast_fp16")];
-            tensor<int32, [4]> var_830 = const()[name = tensor<string, []>("op_830"), val = tensor<int32, [4]>([1, 8, -1, 188])];
-            tensor<fp16, [1, 8, 376, 188]> matrix_bd_21_cast_fp16 = reshape(shape = var_830, x = matrix_bd_19_cast_fp16)[name = tensor<string, []>("matrix_bd_21_cast_fp16")];
-            tensor<int32, [4]> var_834_begin_0 = const()[name = tensor<string, []>("op_834_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
-            tensor<int32, [4]> var_834_end_0 = const()[name = tensor<string, []>("op_834_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
-            tensor<bool, [4]> var_834_end_mask_0 = const()[name = tensor<string, []>("op_834_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
-            tensor<fp16, [1, 8, 375, 188]> var_834_cast_fp16 = slice_by_index(begin = var_834_begin_0, end = var_834_end_0, end_mask = var_834_end_mask_0, x = matrix_bd_21_cast_fp16)[name = tensor<string, []>("op_834_cast_fp16")];
-            tensor<int32, [4]> var_835 = const()[name = tensor<string, []>("op_835"), val = tensor<int32, [4]>([1, 8, 188, 375])];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_23_cast_fp16 = reshape(shape = var_835, x = var_834_cast_fp16)[name = tensor<string, []>("matrix_bd_23_cast_fp16")];
-            tensor<int32, [4]> var_840_begin_0 = const()[name = tensor<string, []>("op_840_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [4]> var_840_end_0 = const()[name = tensor<string, []>("op_840_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
-            tensor<bool, [4]> var_840_end_mask_0 = const()[name = tensor<string, []>("op_840_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
-            tensor<fp16, [1, 8, 188, 188]> var_840_cast_fp16 = slice_by_index(begin = var_840_begin_0, end = var_840_end_0, end_mask = var_840_end_mask_0, x = matrix_bd_23_cast_fp16)[name = tensor<string, []>("op_840_cast_fp16")];
-            tensor<fp16, []> var_841_to_fp16 = const()[name = tensor<string, []>("op_841_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 188, 188]> qk_mask_5_cast_fp16 = mul(x = var_840_cast_fp16, y = var_841_to_fp16)[name = tensor<string, []>("qk_mask_5_cast_fp16")];
-            tensor<int32, [4]> var_845 = const()[name = tensor<string, []>("op_845"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> mh_q_5_cast_fp16 = reshape(shape = var_845, x = query_11_cast_fp16)[name = tensor<string, []>("mh_q_5_cast_fp16")];
-            tensor<fp16, []> var_847_to_fp16 = const()[name = tensor<string, []>("op_847_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 128, 188]> var_848_cast_fp16 = mul(x = mh_q_5_cast_fp16, y = var_847_to_fp16)[name = tensor<string, []>("op_848_cast_fp16")];
-            tensor<int32, [4]> var_851 = const()[name = tensor<string, []>("op_851"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_852_cast_fp16 = reshape(shape = var_851, x = key_5_cast_fp16)[name = tensor<string, []>("op_852_cast_fp16")];
+            tensor<int32, [4]> var_828 = const()[name = tensor<string, []>("op_828"), val = tensor<int32, [4]>([1, 8, -1, 188])];
+            tensor<fp16, [1, 8, 376, 188]> matrix_bd_21_cast_fp16 = reshape(shape = var_828, x = matrix_bd_19_cast_fp16)[name = tensor<string, []>("matrix_bd_21_cast_fp16")];
+            tensor<int32, [4]> var_832_begin_0 = const()[name = tensor<string, []>("op_832_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_832_end_0 = const()[name = tensor<string, []>("op_832_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
+            tensor<bool, [4]> var_832_end_mask_0 = const()[name = tensor<string, []>("op_832_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 375, 188]> var_832_cast_fp16 = slice_by_index(begin = var_832_begin_0, end = var_832_end_0, end_mask = var_832_end_mask_0, x = matrix_bd_21_cast_fp16)[name = tensor<string, []>("op_832_cast_fp16")];
+            tensor<int32, [4]> var_833 = const()[name = tensor<string, []>("op_833"), val = tensor<int32, [4]>([1, 8, 188, 375])];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_23_cast_fp16 = reshape(shape = var_833, x = var_832_cast_fp16)[name = tensor<string, []>("matrix_bd_23_cast_fp16")];
+            tensor<int32, [4]> var_838_begin_0 = const()[name = tensor<string, []>("op_838_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_838_end_0 = const()[name = tensor<string, []>("op_838_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
+            tensor<bool, [4]> var_838_end_mask_0 = const()[name = tensor<string, []>("op_838_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 188, 188]> var_838_cast_fp16 = slice_by_index(begin = var_838_begin_0, end = var_838_end_0, end_mask = var_838_end_mask_0, x = matrix_bd_23_cast_fp16)[name = tensor<string, []>("op_838_cast_fp16")];
+            tensor<fp16, []> var_839_to_fp16 = const()[name = tensor<string, []>("op_839_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 188, 188]> qk_mask_5_cast_fp16 = mul(x = var_838_cast_fp16, y = var_839_to_fp16)[name = tensor<string, []>("qk_mask_5_cast_fp16")];
+            tensor<int32, [4]> var_843 = const()[name = tensor<string, []>("op_843"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> mh_q_5_cast_fp16 = reshape(shape = var_843, x = query_11_cast_fp16)[name = tensor<string, []>("mh_q_5_cast_fp16")];
+            tensor<fp16, []> var_845_to_fp16 = const()[name = tensor<string, []>("op_845_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 128, 188]> var_846_cast_fp16 = mul(x = mh_q_5_cast_fp16, y = var_845_to_fp16)[name = tensor<string, []>("op_846_cast_fp16")];
+            tensor<int32, [4]> var_849 = const()[name = tensor<string, []>("op_849"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_850_cast_fp16 = reshape(shape = var_849, x = key_5_cast_fp16)[name = tensor<string, []>("op_850_cast_fp16")];
             tensor<bool, []> mh_w_9_transpose_x_0 = const()[name = tensor<string, []>("mh_w_9_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> mh_w_9_transpose_y_0 = const()[name = tensor<string, []>("mh_w_9_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 188]> mh_w_9_cast_fp16 = matmul(transpose_x = mh_w_9_transpose_x_0, transpose_y = mh_w_9_transpose_y_0, x = var_848_cast_fp16, y = var_852_cast_fp16)[name = tensor<string, []>("mh_w_9_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> mh_w_9_cast_fp16 = matmul(transpose_x = mh_w_9_transpose_x_0, transpose_y = mh_w_9_transpose_y_0, x = var_846_cast_fp16, y = var_850_cast_fp16)[name = tensor<string, []>("mh_w_9_cast_fp16")];
             tensor<fp16, [1, 8, 188, 188]> mh_w_11_cast_fp16 = add(x = mh_w_9_cast_fp16, y = qk_mask_5_cast_fp16)[name = tensor<string, []>("mh_w_11_cast_fp16")];
-            tensor<fp16, [1, 8, 188, 188]> var_856_cast_fp16 = softmax(axis = var_697, x = mh_w_11_cast_fp16)[name = tensor<string, []>("op_856_cast_fp16")];
-            tensor<int32, [4]> var_857 = const()[name = tensor<string, []>("op_857"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_858_cast_fp16 = reshape(shape = var_857, x = value_5_cast_fp16)[name = tensor<string, []>("op_858_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> var_854_cast_fp16 = softmax(axis = var_695, x = mh_w_11_cast_fp16)[name = tensor<string, []>("op_854_cast_fp16")];
+            tensor<int32, [4]> var_855 = const()[name = tensor<string, []>("op_855"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_856_cast_fp16 = reshape(shape = var_855, x = value_5_cast_fp16)[name = tensor<string, []>("op_856_cast_fp16")];
             tensor<bool, []> attn_5_transpose_x_0 = const()[name = tensor<string, []>("attn_5_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> attn_5_transpose_y_0 = const()[name = tensor<string, []>("attn_5_transpose_y_0"), val = tensor<bool, []>(true)];
-            tensor<fp16, [1, 8, 128, 188]> attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = var_858_cast_fp16, y = var_856_cast_fp16)[name = tensor<string, []>("attn_5_cast_fp16")];
-            tensor<int32, [4]> var_861 = const()[name = tensor<string, []>("op_861"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
-            tensor<fp16, [1, 1024, 1, 188]> input_73_cast_fp16 = reshape(shape = var_861, x = attn_5_cast_fp16)[name = tensor<string, []>("input_73_cast_fp16")];
+            tensor<fp16, [1, 8, 128, 188]> attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = var_856_cast_fp16, y = var_854_cast_fp16)[name = tensor<string, []>("attn_5_cast_fp16")];
+            tensor<int32, [4]> var_859 = const()[name = tensor<string, []>("op_859"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
+            tensor<fp16, [1, 1024, 1, 188]> input_75_cast_fp16 = reshape(shape = var_859, x = attn_5_cast_fp16)[name = tensor<string, []>("input_75_cast_fp16")];
             tensor<string, []> obj_13_pad_type_0 = const()[name = tensor<string, []>("obj_13_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> obj_13_strides_0 = const()[name = tensor<string, []>("obj_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> obj_13_pad_0 = const()[name = tensor<string, []>("obj_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> obj_13_dilations_0 = const()[name = tensor<string, []>("obj_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> obj_13_groups_0 = const()[name = tensor<string, []>("obj_13_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_2_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135386048)))];
-            tensor<fp16, [1, 1024, 1, 188]> obj_13_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = obj_13_dilations_0, groups = obj_13_groups_0, pad = obj_13_pad_0, pad_type = obj_13_pad_type_0, strides = obj_13_strides_0, weight = layers_2_self_attn_o_proj_weight_to_fp16, x = input_73_cast_fp16)[name = tensor<string, []>("obj_13_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_13_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_13_dilations_0, groups = obj_13_groups_0, pad = obj_13_pad_0, pad_type = obj_13_pad_type_0, strides = obj_13_strides_0, weight = layers_2_self_attn_o_proj_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("obj_13_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_25_cast_fp16 = add(x = inputs_23_cast_fp16, y = obj_13_cast_fp16)[name = tensor<string, []>("inputs_25_cast_fp16")];
             tensor<int32, [1]> out_25_axes_0 = const()[name = tensor<string, []>("out_25_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_879_to_fp16 = const()[name = tensor<string, []>("op_879_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_25_cast_fp16 = layer_norm(axes = out_25_axes_0, epsilon = var_879_to_fp16, x = inputs_25_cast_fp16)[name = tensor<string, []>("out_25_cast_fp16")];
-            tensor<fp16, [1024]> input_75_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_75_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137483264)))];
-            tensor<fp16, [1024]> input_75_beta_0_to_fp16 = const()[name = tensor<string, []>("input_75_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137485376)))];
-            tensor<fp16, []> input_75_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_75_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_75_cast_fp16 = batch_norm(beta = input_75_beta_0_to_fp16, epsilon = input_75_epsilon_0_to_fp16, gamma = input_75_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_25_cast_fp16)[name = tensor<string, []>("input_75_cast_fp16")];
-            tensor<string, []> input_77_pad_type_0 = const()[name = tensor<string, []>("input_77_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_77_strides_0 = const()[name = tensor<string, []>("input_77_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_77_pad_0 = const()[name = tensor<string, []>("input_77_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_77_dilations_0 = const()[name = tensor<string, []>("input_77_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_77_groups_0 = const()[name = tensor<string, []>("input_77_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_877_to_fp16 = const()[name = tensor<string, []>("op_877_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_25_cast_fp16 = layer_norm(axes = out_25_axes_0, epsilon = var_877_to_fp16, x = inputs_25_cast_fp16)[name = tensor<string, []>("out_25_cast_fp16")];
+            tensor<fp16, [1024]> input_77_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_77_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137483264)))];
+            tensor<fp16, [1024]> input_77_beta_0_to_fp16 = const()[name = tensor<string, []>("input_77_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137485376)))];
+            tensor<fp16, []> input_77_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_77_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_77_cast_fp16 = batch_norm(beta = input_77_beta_0_to_fp16, epsilon = input_77_epsilon_0_to_fp16, gamma = input_77_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_25_cast_fp16)[name = tensor<string, []>("input_77_cast_fp16")];
+            tensor<string, []> input_79_pad_type_0 = const()[name = tensor<string, []>("input_79_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_79_strides_0 = const()[name = tensor<string, []>("input_79_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_79_pad_0 = const()[name = tensor<string, []>("input_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_79_dilations_0 = const()[name = tensor<string, []>("input_79_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_79_groups_0 = const()[name = tensor<string, []>("input_79_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [2048, 1024, 1, 1]> layers_2_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137487488)))];
-            tensor<fp16, [1, 2048, 1, 188]> input_77_cast_fp16 = conv(dilations = input_77_dilations_0, groups = input_77_groups_0, pad = input_77_pad_0, pad_type = input_77_pad_type_0, strides = input_77_strides_0, weight = layers_2_conv_pointwise_conv1_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("input_77_cast_fp16")];
-            tensor<int32, []> input_79_split_num_splits_0 = const()[name = tensor<string, []>("input_79_split_num_splits_0"), val = tensor<int32, []>(2)];
-            tensor<int32, []> input_79_split_axis_0 = const()[name = tensor<string, []>("input_79_split_axis_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1, 1024, 1, 188]> input_79_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_79_split_cast_fp16_1 = split(axis = input_79_split_axis_0, num_splits = input_79_split_num_splits_0, x = input_77_cast_fp16)[name = tensor<string, []>("input_79_split_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_79_split_1_sigmoid_cast_fp16 = sigmoid(x = input_79_split_cast_fp16_1)[name = tensor<string, []>("input_79_split_1_sigmoid_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_79_cast_fp16 = mul(x = input_79_split_cast_fp16_0, y = input_79_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_79_cast_fp16")];
-            tensor<string, []> input_81_pad_type_0 = const()[name = tensor<string, []>("input_81_pad_type_0"), val = tensor<string, []>("custom")];
-            tensor<int32, [4]> input_81_pad_0 = const()[name = tensor<string, []>("input_81_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
-            tensor<int32, []> input_81_groups_0 = const()[name = tensor<string, []>("input_81_groups_0"), val = tensor<int32, []>(1024)];
-            tensor<int32, [2]> input_81_strides_0 = const()[name = tensor<string, []>("input_81_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> input_81_dilations_0 = const()[name = tensor<string, []>("input_81_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [1, 2048, 1, 188]> input_79_cast_fp16 = conv(dilations = input_79_dilations_0, groups = input_79_groups_0, pad = input_79_pad_0, pad_type = input_79_pad_type_0, strides = input_79_strides_0, weight = layers_2_conv_pointwise_conv1_weight_to_fp16, x = input_77_cast_fp16)[name = tensor<string, []>("input_79_cast_fp16")];
+            tensor<int32, []> input_81_split_num_splits_0 = const()[name = tensor<string, []>("input_81_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> input_81_split_axis_0 = const()[name = tensor<string, []>("input_81_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 1, 188]> input_81_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_81_split_cast_fp16_1 = split(axis = input_81_split_axis_0, num_splits = input_81_split_num_splits_0, x = input_79_cast_fp16)[name = tensor<string, []>("input_81_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_81_split_1_sigmoid_cast_fp16 = sigmoid(x = input_81_split_cast_fp16_1)[name = tensor<string, []>("input_81_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_81_cast_fp16 = mul(x = input_81_split_cast_fp16_0, y = input_81_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_81_cast_fp16")];
+            tensor<string, []> input_83_pad_type_0 = const()[name = tensor<string, []>("input_83_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_83_pad_0 = const()[name = tensor<string, []>("input_83_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
+            tensor<int32, []> input_83_groups_0 = const()[name = tensor<string, []>("input_83_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [2]> input_83_strides_0 = const()[name = tensor<string, []>("input_83_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> input_83_dilations_0 = const()[name = tensor<string, []>("input_83_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<fp16, [1024, 1, 1, 9]> const_272_to_fp16 = const()[name = tensor<string, []>("const_272_to_fp16"), val = tensor<fp16, [1024, 1, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141681856)))];
             tensor<fp16, [1024]> const_273_to_fp16 = const()[name = tensor<string, []>("const_273_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141700352)))];
-            tensor<fp16, [1, 1024, 1, 188]> input_83_cast_fp16 = conv(bias = const_273_to_fp16, dilations = input_81_dilations_0, groups = input_81_groups_0, pad = input_81_pad_0, pad_type = input_81_pad_type_0, strides = input_81_strides_0, weight = const_272_to_fp16, x = input_79_cast_fp16)[name = tensor<string, []>("input_83_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_85_cast_fp16 = silu(x = input_83_cast_fp16)[name = tensor<string, []>("input_85_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_85_cast_fp16 = conv(bias = const_273_to_fp16, dilations = input_83_dilations_0, groups = input_83_groups_0, pad = input_83_pad_0, pad_type = input_83_pad_type_0, strides = input_83_strides_0, weight = const_272_to_fp16, x = input_81_cast_fp16)[name = tensor<string, []>("input_85_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_87_cast_fp16 = silu(x = input_85_cast_fp16)[name = tensor<string, []>("input_87_cast_fp16")];
             tensor<string, []> x_17_pad_type_0 = const()[name = tensor<string, []>("x_17_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_17_strides_0 = const()[name = tensor<string, []>("x_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_17_pad_0 = const()[name = tensor<string, []>("x_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_17_dilations_0 = const()[name = tensor<string, []>("x_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_17_groups_0 = const()[name = tensor<string, []>("x_17_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_2_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141702464)))];
-            tensor<fp16, [1, 1024, 1, 188]> x_17_cast_fp16 = conv(dilations = x_17_dilations_0, groups = x_17_groups_0, pad = x_17_pad_0, pad_type = x_17_pad_type_0, strides = x_17_strides_0, weight = layers_2_conv_pointwise_conv2_weight_to_fp16, x = input_85_cast_fp16)[name = tensor<string, []>("x_17_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> x_17_cast_fp16 = conv(dilations = x_17_dilations_0, groups = x_17_groups_0, pad = x_17_pad_0, pad_type = x_17_pad_type_0, strides = x_17_strides_0, weight = layers_2_conv_pointwise_conv2_weight_to_fp16, x = input_87_cast_fp16)[name = tensor<string, []>("x_17_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_27_cast_fp16 = add(x = inputs_25_cast_fp16, y = x_17_cast_fp16)[name = tensor<string, []>("inputs_27_cast_fp16")];
             tensor<int32, [1]> out_27_axes_0 = const()[name = tensor<string, []>("out_27_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_927_to_fp16 = const()[name = tensor<string, []>("op_927_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_27_cast_fp16 = layer_norm(axes = out_27_axes_0, epsilon = var_927_to_fp16, x = inputs_27_cast_fp16)[name = tensor<string, []>("out_27_cast_fp16")];
-            tensor<fp16, [1024]> input_87_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_87_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143799680)))];
-            tensor<fp16, [1024]> input_87_beta_0_to_fp16 = const()[name = tensor<string, []>("input_87_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143801792)))];
-            tensor<fp16, []> input_87_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_87_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_87_cast_fp16 = batch_norm(beta = input_87_beta_0_to_fp16, epsilon = input_87_epsilon_0_to_fp16, gamma = input_87_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_27_cast_fp16)[name = tensor<string, []>("input_87_cast_fp16")];
-            tensor<string, []> input_89_pad_type_0 = const()[name = tensor<string, []>("input_89_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_89_strides_0 = const()[name = tensor<string, []>("input_89_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_89_pad_0 = const()[name = tensor<string, []>("input_89_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_89_dilations_0 = const()[name = tensor<string, []>("input_89_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_89_groups_0 = const()[name = tensor<string, []>("input_89_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_925_to_fp16 = const()[name = tensor<string, []>("op_925_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_27_cast_fp16 = layer_norm(axes = out_27_axes_0, epsilon = var_925_to_fp16, x = inputs_27_cast_fp16)[name = tensor<string, []>("out_27_cast_fp16")];
+            tensor<fp16, [1024]> input_89_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_89_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143799680)))];
+            tensor<fp16, [1024]> input_89_beta_0_to_fp16 = const()[name = tensor<string, []>("input_89_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143801792)))];
+            tensor<fp16, []> input_89_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_89_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_89_cast_fp16 = batch_norm(beta = input_89_beta_0_to_fp16, epsilon = input_89_epsilon_0_to_fp16, gamma = input_89_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_27_cast_fp16)[name = tensor<string, []>("input_89_cast_fp16")];
+            tensor<string, []> input_91_pad_type_0 = const()[name = tensor<string, []>("input_91_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_91_strides_0 = const()[name = tensor<string, []>("input_91_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_91_pad_0 = const()[name = tensor<string, []>("input_91_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_91_dilations_0 = const()[name = tensor<string, []>("input_91_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_91_groups_0 = const()[name = tensor<string, []>("input_91_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_2_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_feed_forward2_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143803904)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_89_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_89_dilations_0, groups = input_89_groups_0, pad = input_89_pad_0, pad_type = input_89_pad_type_0, strides = input_89_strides_0, weight = layers_2_feed_forward2_fc1_weight_to_fp16, x = input_87_cast_fp16)[name = tensor<string, []>("input_89_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_91_cast_fp16 = silu(x = input_89_cast_fp16)[name = tensor<string, []>("input_91_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_91_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_91_dilations_0, groups = input_91_groups_0, pad = input_91_pad_0, pad_type = input_91_pad_type_0, strides = input_91_strides_0, weight = layers_2_feed_forward2_fc1_weight_to_fp16, x = input_89_cast_fp16)[name = tensor<string, []>("input_91_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_93_cast_fp16 = silu(x = input_91_cast_fp16)[name = tensor<string, []>("input_93_cast_fp16")];
             tensor<string, []> x_19_pad_type_0 = const()[name = tensor<string, []>("x_19_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_19_strides_0 = const()[name = tensor<string, []>("x_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_19_pad_0 = const()[name = tensor<string, []>("x_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_19_dilations_0 = const()[name = tensor<string, []>("x_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_19_groups_0 = const()[name = tensor<string, []>("x_19_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_955_weight_0_to_fp16 = const()[name = tensor<string, []>("op_955_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(152192576)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_955_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_19_dilations_0, groups = x_19_groups_0, pad = x_19_pad_0, pad_type = x_19_pad_type_0, strides = x_19_strides_0, weight = var_955_weight_0_to_fp16, x = input_91_cast_fp16)[name = tensor<string, []>("op_955_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_29_cast_fp16 = add(x = inputs_27_cast_fp16, y = var_955_cast_fp16)[name = tensor<string, []>("inputs_29_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_953_weight_0_to_fp16 = const()[name = tensor<string, []>("op_953_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(152192576)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_953_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_19_dilations_0, groups = x_19_groups_0, pad = x_19_pad_0, pad_type = x_19_pad_type_0, strides = x_19_strides_0, weight = var_953_weight_0_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("op_953_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_29_cast_fp16 = add(x = inputs_27_cast_fp16, y = var_953_cast_fp16)[name = tensor<string, []>("inputs_29_cast_fp16")];
             tensor<int32, [1]> out_29_axes_0 = const()[name = tensor<string, []>("out_29_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_965_to_fp16 = const()[name = tensor<string, []>("op_965_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_29_cast_fp16 = layer_norm(axes = out_29_axes_0, epsilon = var_965_to_fp16, x = inputs_29_cast_fp16)[name = tensor<string, []>("out_29_cast_fp16")];
+            tensor<fp16, []> var_963_to_fp16 = const()[name = tensor<string, []>("op_963_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_29_cast_fp16 = layer_norm(axes = out_29_axes_0, epsilon = var_963_to_fp16, x = inputs_29_cast_fp16)[name = tensor<string, []>("out_29_cast_fp16")];
             tensor<fp16, [1024]> inputs_31_gamma_0_to_fp16 = const()[name = tensor<string, []>("inputs_31_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160581248)))];
             tensor<fp16, [1024]> inputs_31_beta_0_to_fp16 = const()[name = tensor<string, []>("inputs_31_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160583360)))];
             tensor<fp16, []> inputs_31_epsilon_0_to_fp16 = const()[name = tensor<string, []>("inputs_31_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_31_cast_fp16 = batch_norm(beta = inputs_31_beta_0_to_fp16, epsilon = inputs_31_epsilon_0_to_fp16, gamma = inputs_31_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_29_cast_fp16)[name = tensor<string, []>("inputs_31_cast_fp16")];
-            tensor<int32, []> var_979 = const()[name = tensor<string, []>("op_979"), val = tensor<int32, []>(3)];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_31_cast_fp16 = batch_norm(beta = inputs_31_beta_0_to_fp16, epsilon = inputs_31_epsilon_0_to_fp16, gamma = inputs_31_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_29_cast_fp16)[name = tensor<string, []>("inputs_31_cast_fp16")];
+            tensor<int32, []> var_977 = const()[name = tensor<string, []>("op_977"), val = tensor<int32, []>(3)];
             tensor<int32, [1]> out_31_axes_0 = const()[name = tensor<string, []>("out_31_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_1010_to_fp16 = const()[name = tensor<string, []>("op_1010_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_31_cast_fp16 = layer_norm(axes = out_31_axes_0, epsilon = var_1010_to_fp16, x = inputs_31_cast_fp16)[name = tensor<string, []>("out_31_cast_fp16")];
-            tensor<fp16, [1024]> input_93_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_93_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160585472)))];
-            tensor<fp16, [1024]> input_93_beta_0_to_fp16 = const()[name = tensor<string, []>("input_93_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160587584)))];
-            tensor<fp16, []> input_93_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_93_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_93_cast_fp16 = batch_norm(beta = input_93_beta_0_to_fp16, epsilon = input_93_epsilon_0_to_fp16, gamma = input_93_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_31_cast_fp16)[name = tensor<string, []>("input_93_cast_fp16")];
-            tensor<string, []> input_95_pad_type_0 = const()[name = tensor<string, []>("input_95_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_95_strides_0 = const()[name = tensor<string, []>("input_95_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_95_pad_0 = const()[name = tensor<string, []>("input_95_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_95_dilations_0 = const()[name = tensor<string, []>("input_95_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_95_groups_0 = const()[name = tensor<string, []>("input_95_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_1008_to_fp16 = const()[name = tensor<string, []>("op_1008_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_31_cast_fp16 = layer_norm(axes = out_31_axes_0, epsilon = var_1008_to_fp16, x = inputs_31_cast_fp16)[name = tensor<string, []>("out_31_cast_fp16")];
+            tensor<fp16, [1024]> input_95_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_95_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160585472)))];
+            tensor<fp16, [1024]> input_95_beta_0_to_fp16 = const()[name = tensor<string, []>("input_95_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160587584)))];
+            tensor<fp16, []> input_95_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_95_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_95_cast_fp16 = batch_norm(beta = input_95_beta_0_to_fp16, epsilon = input_95_epsilon_0_to_fp16, gamma = input_95_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_31_cast_fp16)[name = tensor<string, []>("input_95_cast_fp16")];
+            tensor<string, []> input_97_pad_type_0 = const()[name = tensor<string, []>("input_97_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_97_strides_0 = const()[name = tensor<string, []>("input_97_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_97_pad_0 = const()[name = tensor<string, []>("input_97_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_97_dilations_0 = const()[name = tensor<string, []>("input_97_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_97_groups_0 = const()[name = tensor<string, []>("input_97_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_3_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_feed_forward1_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160589696)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_95_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_95_dilations_0, groups = input_95_groups_0, pad = input_95_pad_0, pad_type = input_95_pad_type_0, strides = input_95_strides_0, weight = layers_3_feed_forward1_fc1_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("input_95_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_97_cast_fp16 = silu(x = input_95_cast_fp16)[name = tensor<string, []>("input_97_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_97_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_97_dilations_0, groups = input_97_groups_0, pad = input_97_pad_0, pad_type = input_97_pad_type_0, strides = input_97_strides_0, weight = layers_3_feed_forward1_fc1_weight_to_fp16, x = input_95_cast_fp16)[name = tensor<string, []>("input_97_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_99_cast_fp16 = silu(x = input_97_cast_fp16)[name = tensor<string, []>("input_99_cast_fp16")];
             tensor<string, []> x_21_pad_type_0 = const()[name = tensor<string, []>("x_21_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_21_strides_0 = const()[name = tensor<string, []>("x_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_21_pad_0 = const()[name = tensor<string, []>("x_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_21_dilations_0 = const()[name = tensor<string, []>("x_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_21_groups_0 = const()[name = tensor<string, []>("x_21_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_1038_weight_0_to_fp16 = const()[name = tensor<string, []>("op_1038_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168978368)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_1038_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_21_dilations_0, groups = x_21_groups_0, pad = x_21_pad_0, pad_type = x_21_pad_type_0, strides = x_21_strides_0, weight = var_1038_weight_0_to_fp16, x = input_97_cast_fp16)[name = tensor<string, []>("op_1038_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_33_cast_fp16 = add(x = inputs_31_cast_fp16, y = var_1038_cast_fp16)[name = tensor<string, []>("inputs_33_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_1036_weight_0_to_fp16 = const()[name = tensor<string, []>("op_1036_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168978368)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_1036_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_21_dilations_0, groups = x_21_groups_0, pad = x_21_pad_0, pad_type = x_21_pad_type_0, strides = x_21_strides_0, weight = var_1036_weight_0_to_fp16, x = input_99_cast_fp16)[name = tensor<string, []>("op_1036_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_33_cast_fp16 = add(x = inputs_31_cast_fp16, y = var_1036_cast_fp16)[name = tensor<string, []>("inputs_33_cast_fp16")];
             tensor<int32, [1]> out_33_axes_0 = const()[name = tensor<string, []>("out_33_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_1048_to_fp16 = const()[name = tensor<string, []>("op_1048_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_33_cast_fp16 = layer_norm(axes = out_33_axes_0, epsilon = var_1048_to_fp16, x = inputs_33_cast_fp16)[name = tensor<string, []>("out_33_cast_fp16")];
+            tensor<fp16, []> var_1046_to_fp16 = const()[name = tensor<string, []>("op_1046_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_33_cast_fp16 = layer_norm(axes = out_33_axes_0, epsilon = var_1046_to_fp16, x = inputs_33_cast_fp16)[name = tensor<string, []>("out_33_cast_fp16")];
             tensor<fp16, [1024]> obj_15_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_15_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(177367040)))];
             tensor<fp16, [1024]> obj_15_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_15_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(177369152)))];
             tensor<fp16, []> obj_15_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_15_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> obj_15_cast_fp16 = batch_norm(beta = obj_15_beta_0_to_fp16, epsilon = obj_15_epsilon_0_to_fp16, gamma = obj_15_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_33_cast_fp16)[name = tensor<string, []>("obj_15_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_15_cast_fp16 = batch_norm(beta = obj_15_beta_0_to_fp16, epsilon = obj_15_epsilon_0_to_fp16, gamma = obj_15_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_33_cast_fp16)[name = tensor<string, []>("obj_15_cast_fp16")];
             tensor<string, []> query_13_pad_type_0 = const()[name = tensor<string, []>("query_13_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> query_13_strides_0 = const()[name = tensor<string, []>("query_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> query_13_pad_0 = const()[name = tensor<string, []>("query_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> query_13_dilations_0 = const()[name = tensor<string, []>("query_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> query_13_groups_0 = const()[name = tensor<string, []>("query_13_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_3_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(177371264)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_13_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = query_13_dilations_0, groups = query_13_groups_0, pad = query_13_pad_0, pad_type = query_13_pad_type_0, strides = query_13_strides_0, weight = layers_3_self_attn_q_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor<string, []>("query_13_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> query_13_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_13_dilations_0, groups = query_13_groups_0, pad = query_13_pad_0, pad_type = query_13_pad_type_0, strides = query_13_strides_0, weight = layers_3_self_attn_q_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor<string, []>("query_13_cast_fp16")];
             tensor<string, []> key_7_pad_type_0 = const()[name = tensor<string, []>("key_7_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> key_7_strides_0 = const()[name = tensor<string, []>("key_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> key_7_pad_0 = const()[name = tensor<string, []>("key_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
@@ -652,173 +654,173 @@ program(1.0)
             tensor<int32, [2]> value_7_dilations_0 = const()[name = tensor<string, []>("value_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> value_7_groups_0 = const()[name = tensor<string, []>("value_7_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_3_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181565696)))];
-            tensor<fp16, [1, 1024, 1, 188]> value_7_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = value_7_dilations_0, groups = value_7_groups_0, pad = value_7_pad_0, pad_type = value_7_pad_type_0, strides = value_7_strides_0, weight = layers_3_self_attn_v_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor<string, []>("value_7_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_1086_to_fp16 = const()[name = tensor<string, []>("op_1086_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183662912)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_15_cast_fp16 = add(x = query_13_cast_fp16, y = var_1086_to_fp16)[name = tensor<string, []>("query_15_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_1089_to_fp16 = const()[name = tensor<string, []>("op_1089_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183665024)))];
-            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_7_cast_fp16 = add(x = query_13_cast_fp16, y = var_1089_to_fp16)[name = tensor<string, []>("q_with_bias_v_7_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> value_7_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_7_dilations_0, groups = value_7_groups_0, pad = value_7_pad_0, pad_type = value_7_pad_type_0, strides = value_7_strides_0, weight = layers_3_self_attn_v_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor<string, []>("value_7_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_1084_to_fp16 = const()[name = tensor<string, []>("op_1084_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183662912)))];
+            tensor<fp16, [1, 1024, 1, 188]> query_15_cast_fp16 = add(x = query_13_cast_fp16, y = var_1084_to_fp16)[name = tensor<string, []>("query_15_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_1087_to_fp16 = const()[name = tensor<string, []>("op_1087_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183665024)))];
+            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_7_cast_fp16 = add(x = query_13_cast_fp16, y = var_1087_to_fp16)[name = tensor<string, []>("q_with_bias_v_7_cast_fp16")];
             tensor<string, []> p_7_pad_type_0 = const()[name = tensor<string, []>("p_7_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> p_7_strides_0 = const()[name = tensor<string, []>("p_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> p_7_pad_0 = const()[name = tensor<string, []>("p_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> p_7_dilations_0 = const()[name = tensor<string, []>("p_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> p_7_groups_0 = const()[name = tensor<string, []>("p_7_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_3_self_attn_linear_pos_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_linear_pos_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183667136)))];
-            tensor<fp16, [1, 1024, 1, 375]> p_7_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = p_7_dilations_0, groups = p_7_groups_0, pad = p_7_pad_0, pad_type = p_7_pad_type_0, strides = p_7_strides_0, weight = layers_3_self_attn_linear_pos_weight_to_fp16, x = pos_enc_to_fp16)[name = tensor<string, []>("p_7_cast_fp16")];
-            tensor<int32, [4]> var_1100 = const()[name = tensor<string, []>("op_1100"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_1101_cast_fp16 = reshape(shape = var_1100, x = q_with_bias_v_7_cast_fp16)[name = tensor<string, []>("op_1101_cast_fp16")];
-            tensor<int32, [4]> var_1102 = const()[name = tensor<string, []>("op_1102"), val = tensor<int32, [4]>([1, 8, 128, -1])];
-            tensor<fp16, [1, 8, 128, 375]> var_1103_cast_fp16 = reshape(shape = var_1102, x = p_7_cast_fp16)[name = tensor<string, []>("op_1103_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 375]> p_7_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_7_dilations_0, groups = p_7_groups_0, pad = p_7_pad_0, pad_type = p_7_pad_type_0, strides = p_7_strides_0, weight = layers_3_self_attn_linear_pos_weight_to_fp16, x = obj_3_cast_fp16)[name = tensor<string, []>("p_7_cast_fp16")];
+            tensor<int32, [4]> var_1098 = const()[name = tensor<string, []>("op_1098"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_1099_cast_fp16 = reshape(shape = var_1098, x = q_with_bias_v_7_cast_fp16)[name = tensor<string, []>("op_1099_cast_fp16")];
+            tensor<int32, [4]> var_1100 = const()[name = tensor<string, []>("op_1100"), val = tensor<int32, [4]>([1, 8, 128, -1])];
+            tensor<fp16, [1, 8, 128, 375]> var_1101_cast_fp16 = reshape(shape = var_1100, x = p_7_cast_fp16)[name = tensor<string, []>("op_1101_cast_fp16")];
             tensor<bool, []> matrix_bd_25_transpose_x_0 = const()[name = tensor<string, []>("matrix_bd_25_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> matrix_bd_25_transpose_y_0 = const()[name = tensor<string, []>("matrix_bd_25_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_25_cast_fp16 = matmul(transpose_x = matrix_bd_25_transpose_x_0, transpose_y = matrix_bd_25_transpose_y_0, x = var_1101_cast_fp16, y = var_1103_cast_fp16)[name = tensor<string, []>("matrix_bd_25_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_25_cast_fp16 = matmul(transpose_x = matrix_bd_25_transpose_x_0, transpose_y = matrix_bd_25_transpose_y_0, x = var_1099_cast_fp16, y = var_1101_cast_fp16)[name = tensor<string, []>("matrix_bd_25_cast_fp16")];
             tensor<int32, [8]> matrix_bd_27_pad_0 = const()[name = tensor<string, []>("matrix_bd_27_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
             tensor<string, []> matrix_bd_27_mode_0 = const()[name = tensor<string, []>("matrix_bd_27_mode_0"), val = tensor<string, []>("constant")];
             tensor<fp16, []> const_43_to_fp16 = const()[name = tensor<string, []>("const_43_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
             tensor<fp16, [1, 8, 188, 376]> matrix_bd_27_cast_fp16 = pad(constant_val = const_43_to_fp16, mode = matrix_bd_27_mode_0, pad = matrix_bd_27_pad_0, x = matrix_bd_25_cast_fp16)[name = tensor<string, []>("matrix_bd_27_cast_fp16")];
-            tensor<int32, [4]> var_1112 = const()[name = tensor<string, []>("op_1112"), val = tensor<int32, [4]>([1, 8, -1, 188])];
-            tensor<fp16, [1, 8, 376, 188]> matrix_bd_29_cast_fp16 = reshape(shape = var_1112, x = matrix_bd_27_cast_fp16)[name = tensor<string, []>("matrix_bd_29_cast_fp16")];
-            tensor<int32, [4]> var_1116_begin_0 = const()[name = tensor<string, []>("op_1116_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
-            tensor<int32, [4]> var_1116_end_0 = const()[name = tensor<string, []>("op_1116_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
-            tensor<bool, [4]> var_1116_end_mask_0 = const()[name = tensor<string, []>("op_1116_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
-            tensor<fp16, [1, 8, 375, 188]> var_1116_cast_fp16 = slice_by_index(begin = var_1116_begin_0, end = var_1116_end_0, end_mask = var_1116_end_mask_0, x = matrix_bd_29_cast_fp16)[name = tensor<string, []>("op_1116_cast_fp16")];
-            tensor<int32, [4]> var_1117 = const()[name = tensor<string, []>("op_1117"), val = tensor<int32, [4]>([1, 8, 188, 375])];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_31_cast_fp16 = reshape(shape = var_1117, x = var_1116_cast_fp16)[name = tensor<string, []>("matrix_bd_31_cast_fp16")];
-            tensor<int32, [4]> var_1122_begin_0 = const()[name = tensor<string, []>("op_1122_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [4]> var_1122_end_0 = const()[name = tensor<string, []>("op_1122_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
-            tensor<bool, [4]> var_1122_end_mask_0 = const()[name = tensor<string, []>("op_1122_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
-            tensor<fp16, [1, 8, 188, 188]> var_1122_cast_fp16 = slice_by_index(begin = var_1122_begin_0, end = var_1122_end_0, end_mask = var_1122_end_mask_0, x = matrix_bd_31_cast_fp16)[name = tensor<string, []>("op_1122_cast_fp16")];
-            tensor<fp16, []> var_1123_to_fp16 = const()[name = tensor<string, []>("op_1123_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 188, 188]> qk_mask_7_cast_fp16 = mul(x = var_1122_cast_fp16, y = var_1123_to_fp16)[name = tensor<string, []>("qk_mask_7_cast_fp16")];
-            tensor<int32, [4]> var_1127 = const()[name = tensor<string, []>("op_1127"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> mh_q_7_cast_fp16 = reshape(shape = var_1127, x = query_15_cast_fp16)[name = tensor<string, []>("mh_q_7_cast_fp16")];
-            tensor<fp16, []> var_1129_to_fp16 = const()[name = tensor<string, []>("op_1129_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 128, 188]> var_1130_cast_fp16 = mul(x = mh_q_7_cast_fp16, y = var_1129_to_fp16)[name = tensor<string, []>("op_1130_cast_fp16")];
-            tensor<int32, [4]> var_1133 = const()[name = tensor<string, []>("op_1133"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_1134_cast_fp16 = reshape(shape = var_1133, x = key_7_cast_fp16)[name = tensor<string, []>("op_1134_cast_fp16")];
+            tensor<int32, [4]> var_1110 = const()[name = tensor<string, []>("op_1110"), val = tensor<int32, [4]>([1, 8, -1, 188])];
+            tensor<fp16, [1, 8, 376, 188]> matrix_bd_29_cast_fp16 = reshape(shape = var_1110, x = matrix_bd_27_cast_fp16)[name = tensor<string, []>("matrix_bd_29_cast_fp16")];
+            tensor<int32, [4]> var_1114_begin_0 = const()[name = tensor<string, []>("op_1114_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1114_end_0 = const()[name = tensor<string, []>("op_1114_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
+            tensor<bool, [4]> var_1114_end_mask_0 = const()[name = tensor<string, []>("op_1114_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 375, 188]> var_1114_cast_fp16 = slice_by_index(begin = var_1114_begin_0, end = var_1114_end_0, end_mask = var_1114_end_mask_0, x = matrix_bd_29_cast_fp16)[name = tensor<string, []>("op_1114_cast_fp16")];
+            tensor<int32, [4]> var_1115 = const()[name = tensor<string, []>("op_1115"), val = tensor<int32, [4]>([1, 8, 188, 375])];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_31_cast_fp16 = reshape(shape = var_1115, x = var_1114_cast_fp16)[name = tensor<string, []>("matrix_bd_31_cast_fp16")];
+            tensor<int32, [4]> var_1120_begin_0 = const()[name = tensor<string, []>("op_1120_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_1120_end_0 = const()[name = tensor<string, []>("op_1120_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
+            tensor<bool, [4]> var_1120_end_mask_0 = const()[name = tensor<string, []>("op_1120_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 188, 188]> var_1120_cast_fp16 = slice_by_index(begin = var_1120_begin_0, end = var_1120_end_0, end_mask = var_1120_end_mask_0, x = matrix_bd_31_cast_fp16)[name = tensor<string, []>("op_1120_cast_fp16")];
+            tensor<fp16, []> var_1121_to_fp16 = const()[name = tensor<string, []>("op_1121_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 188, 188]> qk_mask_7_cast_fp16 = mul(x = var_1120_cast_fp16, y = var_1121_to_fp16)[name = tensor<string, []>("qk_mask_7_cast_fp16")];
+            tensor<int32, [4]> var_1125 = const()[name = tensor<string, []>("op_1125"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> mh_q_7_cast_fp16 = reshape(shape = var_1125, x = query_15_cast_fp16)[name = tensor<string, []>("mh_q_7_cast_fp16")];
+            tensor<fp16, []> var_1127_to_fp16 = const()[name = tensor<string, []>("op_1127_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 128, 188]> var_1128_cast_fp16 = mul(x = mh_q_7_cast_fp16, y = var_1127_to_fp16)[name = tensor<string, []>("op_1128_cast_fp16")];
+            tensor<int32, [4]> var_1131 = const()[name = tensor<string, []>("op_1131"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_1132_cast_fp16 = reshape(shape = var_1131, x = key_7_cast_fp16)[name = tensor<string, []>("op_1132_cast_fp16")];
             tensor<bool, []> mh_w_13_transpose_x_0 = const()[name = tensor<string, []>("mh_w_13_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> mh_w_13_transpose_y_0 = const()[name = tensor<string, []>("mh_w_13_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 188]> mh_w_13_cast_fp16 = matmul(transpose_x = mh_w_13_transpose_x_0, transpose_y = mh_w_13_transpose_y_0, x = var_1130_cast_fp16, y = var_1134_cast_fp16)[name = tensor<string, []>("mh_w_13_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> mh_w_13_cast_fp16 = matmul(transpose_x = mh_w_13_transpose_x_0, transpose_y = mh_w_13_transpose_y_0, x = var_1128_cast_fp16, y = var_1132_cast_fp16)[name = tensor<string, []>("mh_w_13_cast_fp16")];
             tensor<fp16, [1, 8, 188, 188]> mh_w_15_cast_fp16 = add(x = mh_w_13_cast_fp16, y = qk_mask_7_cast_fp16)[name = tensor<string, []>("mh_w_15_cast_fp16")];
-            tensor<fp16, [1, 8, 188, 188]> var_1138_cast_fp16 = softmax(axis = var_979, x = mh_w_15_cast_fp16)[name = tensor<string, []>("op_1138_cast_fp16")];
-            tensor<int32, [4]> var_1139 = const()[name = tensor<string, []>("op_1139"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_1140_cast_fp16 = reshape(shape = var_1139, x = value_7_cast_fp16)[name = tensor<string, []>("op_1140_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> var_1136_cast_fp16 = softmax(axis = var_977, x = mh_w_15_cast_fp16)[name = tensor<string, []>("op_1136_cast_fp16")];
+            tensor<int32, [4]> var_1137 = const()[name = tensor<string, []>("op_1137"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_1138_cast_fp16 = reshape(shape = var_1137, x = value_7_cast_fp16)[name = tensor<string, []>("op_1138_cast_fp16")];
             tensor<bool, []> attn_7_transpose_x_0 = const()[name = tensor<string, []>("attn_7_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> attn_7_transpose_y_0 = const()[name = tensor<string, []>("attn_7_transpose_y_0"), val = tensor<bool, []>(true)];
-            tensor<fp16, [1, 8, 128, 188]> attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = var_1140_cast_fp16, y = var_1138_cast_fp16)[name = tensor<string, []>("attn_7_cast_fp16")];
-            tensor<int32, [4]> var_1143 = const()[name = tensor<string, []>("op_1143"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
-            tensor<fp16, [1, 1024, 1, 188]> input_99_cast_fp16 = reshape(shape = var_1143, x = attn_7_cast_fp16)[name = tensor<string, []>("input_99_cast_fp16")];
+            tensor<fp16, [1, 8, 128, 188]> attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = var_1138_cast_fp16, y = var_1136_cast_fp16)[name = tensor<string, []>("attn_7_cast_fp16")];
+            tensor<int32, [4]> var_1141 = const()[name = tensor<string, []>("op_1141"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
+            tensor<fp16, [1, 1024, 1, 188]> input_101_cast_fp16 = reshape(shape = var_1141, x = attn_7_cast_fp16)[name = tensor<string, []>("input_101_cast_fp16")];
             tensor<string, []> obj_17_pad_type_0 = const()[name = tensor<string, []>("obj_17_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> obj_17_strides_0 = const()[name = tensor<string, []>("obj_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> obj_17_pad_0 = const()[name = tensor<string, []>("obj_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> obj_17_dilations_0 = const()[name = tensor<string, []>("obj_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> obj_17_groups_0 = const()[name = tensor<string, []>("obj_17_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_3_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(185764352)))];
-            tensor<fp16, [1, 1024, 1, 188]> obj_17_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = obj_17_dilations_0, groups = obj_17_groups_0, pad = obj_17_pad_0, pad_type = obj_17_pad_type_0, strides = obj_17_strides_0, weight = layers_3_self_attn_o_proj_weight_to_fp16, x = input_99_cast_fp16)[name = tensor<string, []>("obj_17_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_17_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_17_dilations_0, groups = obj_17_groups_0, pad = obj_17_pad_0, pad_type = obj_17_pad_type_0, strides = obj_17_strides_0, weight = layers_3_self_attn_o_proj_weight_to_fp16, x = input_101_cast_fp16)[name = tensor<string, []>("obj_17_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_35_cast_fp16 = add(x = inputs_33_cast_fp16, y = obj_17_cast_fp16)[name = tensor<string, []>("inputs_35_cast_fp16")];
             tensor<int32, [1]> out_35_axes_0 = const()[name = tensor<string, []>("out_35_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_1161_to_fp16 = const()[name = tensor<string, []>("op_1161_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_35_cast_fp16 = layer_norm(axes = out_35_axes_0, epsilon = var_1161_to_fp16, x = inputs_35_cast_fp16)[name = tensor<string, []>("out_35_cast_fp16")];
-            tensor<fp16, [1024]> input_101_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_101_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187861568)))];
-            tensor<fp16, [1024]> input_101_beta_0_to_fp16 = const()[name = tensor<string, []>("input_101_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187863680)))];
-            tensor<fp16, []> input_101_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_101_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_101_cast_fp16 = batch_norm(beta = input_101_beta_0_to_fp16, epsilon = input_101_epsilon_0_to_fp16, gamma = input_101_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_35_cast_fp16)[name = tensor<string, []>("input_101_cast_fp16")];
-            tensor<string, []> input_103_pad_type_0 = const()[name = tensor<string, []>("input_103_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_103_strides_0 = const()[name = tensor<string, []>("input_103_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_103_pad_0 = const()[name = tensor<string, []>("input_103_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_103_dilations_0 = const()[name = tensor<string, []>("input_103_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_103_groups_0 = const()[name = tensor<string, []>("input_103_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_1159_to_fp16 = const()[name = tensor<string, []>("op_1159_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_35_cast_fp16 = layer_norm(axes = out_35_axes_0, epsilon = var_1159_to_fp16, x = inputs_35_cast_fp16)[name = tensor<string, []>("out_35_cast_fp16")];
+            tensor<fp16, [1024]> input_103_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_103_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187861568)))];
+            tensor<fp16, [1024]> input_103_beta_0_to_fp16 = const()[name = tensor<string, []>("input_103_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187863680)))];
+            tensor<fp16, []> input_103_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_103_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_103_cast_fp16 = batch_norm(beta = input_103_beta_0_to_fp16, epsilon = input_103_epsilon_0_to_fp16, gamma = input_103_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_35_cast_fp16)[name = tensor<string, []>("input_103_cast_fp16")];
+            tensor<string, []> input_105_pad_type_0 = const()[name = tensor<string, []>("input_105_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_105_strides_0 = const()[name = tensor<string, []>("input_105_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_105_pad_0 = const()[name = tensor<string, []>("input_105_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_105_dilations_0 = const()[name = tensor<string, []>("input_105_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_105_groups_0 = const()[name = tensor<string, []>("input_105_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [2048, 1024, 1, 1]> layers_3_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187865792)))];
-            tensor<fp16, [1, 2048, 1, 188]> input_103_cast_fp16 = conv(dilations = input_103_dilations_0, groups = input_103_groups_0, pad = input_103_pad_0, pad_type = input_103_pad_type_0, strides = input_103_strides_0, weight = layers_3_conv_pointwise_conv1_weight_to_fp16, x = input_101_cast_fp16)[name = tensor<string, []>("input_103_cast_fp16")];
-            tensor<int32, []> input_105_split_num_splits_0 = const()[name = tensor<string, []>("input_105_split_num_splits_0"), val = tensor<int32, []>(2)];
-            tensor<int32, []> input_105_split_axis_0 = const()[name = tensor<string, []>("input_105_split_axis_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1, 1024, 1, 188]> input_105_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_105_split_cast_fp16_1 = split(axis = input_105_split_axis_0, num_splits = input_105_split_num_splits_0, x = input_103_cast_fp16)[name = tensor<string, []>("input_105_split_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_105_split_1_sigmoid_cast_fp16 = sigmoid(x = input_105_split_cast_fp16_1)[name = tensor<string, []>("input_105_split_1_sigmoid_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_105_cast_fp16 = mul(x = input_105_split_cast_fp16_0, y = input_105_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_105_cast_fp16")];
-            tensor<string, []> input_107_pad_type_0 = const()[name = tensor<string, []>("input_107_pad_type_0"), val = tensor<string, []>("custom")];
-            tensor<int32, [4]> input_107_pad_0 = const()[name = tensor<string, []>("input_107_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
-            tensor<int32, []> input_107_groups_0 = const()[name = tensor<string, []>("input_107_groups_0"), val = tensor<int32, []>(1024)];
-            tensor<int32, [2]> input_107_strides_0 = const()[name = tensor<string, []>("input_107_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> input_107_dilations_0 = const()[name = tensor<string, []>("input_107_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [1, 2048, 1, 188]> input_105_cast_fp16 = conv(dilations = input_105_dilations_0, groups = input_105_groups_0, pad = input_105_pad_0, pad_type = input_105_pad_type_0, strides = input_105_strides_0, weight = layers_3_conv_pointwise_conv1_weight_to_fp16, x = input_103_cast_fp16)[name = tensor<string, []>("input_105_cast_fp16")];
+            tensor<int32, []> input_107_split_num_splits_0 = const()[name = tensor<string, []>("input_107_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> input_107_split_axis_0 = const()[name = tensor<string, []>("input_107_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 1, 188]> input_107_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_107_split_cast_fp16_1 = split(axis = input_107_split_axis_0, num_splits = input_107_split_num_splits_0, x = input_105_cast_fp16)[name = tensor<string, []>("input_107_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_107_split_1_sigmoid_cast_fp16 = sigmoid(x = input_107_split_cast_fp16_1)[name = tensor<string, []>("input_107_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_107_cast_fp16 = mul(x = input_107_split_cast_fp16_0, y = input_107_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_107_cast_fp16")];
+            tensor<string, []> input_109_pad_type_0 = const()[name = tensor<string, []>("input_109_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_109_pad_0 = const()[name = tensor<string, []>("input_109_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
+            tensor<int32, []> input_109_groups_0 = const()[name = tensor<string, []>("input_109_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [2]> input_109_strides_0 = const()[name = tensor<string, []>("input_109_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> input_109_dilations_0 = const()[name = tensor<string, []>("input_109_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<fp16, [1024, 1, 1, 9]> const_274_to_fp16 = const()[name = tensor<string, []>("const_274_to_fp16"), val = tensor<fp16, [1024, 1, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192060160)))];
             tensor<fp16, [1024]> const_275_to_fp16 = const()[name = tensor<string, []>("const_275_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192078656)))];
-            tensor<fp16, [1, 1024, 1, 188]> input_109_cast_fp16 = conv(bias = const_275_to_fp16, dilations = input_107_dilations_0, groups = input_107_groups_0, pad = input_107_pad_0, pad_type = input_107_pad_type_0, strides = input_107_strides_0, weight = const_274_to_fp16, x = input_105_cast_fp16)[name = tensor<string, []>("input_109_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_111_cast_fp16 = silu(x = input_109_cast_fp16)[name = tensor<string, []>("input_111_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_111_cast_fp16 = conv(bias = const_275_to_fp16, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = const_274_to_fp16, x = input_107_cast_fp16)[name = tensor<string, []>("input_111_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_113_cast_fp16 = silu(x = input_111_cast_fp16)[name = tensor<string, []>("input_113_cast_fp16")];
             tensor<string, []> x_23_pad_type_0 = const()[name = tensor<string, []>("x_23_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_23_strides_0 = const()[name = tensor<string, []>("x_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_23_pad_0 = const()[name = tensor<string, []>("x_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_23_dilations_0 = const()[name = tensor<string, []>("x_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_23_groups_0 = const()[name = tensor<string, []>("x_23_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_3_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192080768)))];
-            tensor<fp16, [1, 1024, 1, 188]> x_23_cast_fp16 = conv(dilations = x_23_dilations_0, groups = x_23_groups_0, pad = x_23_pad_0, pad_type = x_23_pad_type_0, strides = x_23_strides_0, weight = layers_3_conv_pointwise_conv2_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("x_23_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> x_23_cast_fp16 = conv(dilations = x_23_dilations_0, groups = x_23_groups_0, pad = x_23_pad_0, pad_type = x_23_pad_type_0, strides = x_23_strides_0, weight = layers_3_conv_pointwise_conv2_weight_to_fp16, x = input_113_cast_fp16)[name = tensor<string, []>("x_23_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_37_cast_fp16 = add(x = inputs_35_cast_fp16, y = x_23_cast_fp16)[name = tensor<string, []>("inputs_37_cast_fp16")];
             tensor<int32, [1]> out_37_axes_0 = const()[name = tensor<string, []>("out_37_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_1209_to_fp16 = const()[name = tensor<string, []>("op_1209_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_37_cast_fp16 = layer_norm(axes = out_37_axes_0, epsilon = var_1209_to_fp16, x = inputs_37_cast_fp16)[name = tensor<string, []>("out_37_cast_fp16")];
-            tensor<fp16, [1024]> input_113_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_113_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194177984)))];
-            tensor<fp16, [1024]> input_113_beta_0_to_fp16 = const()[name = tensor<string, []>("input_113_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194180096)))];
-            tensor<fp16, []> input_113_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_113_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_113_cast_fp16 = batch_norm(beta = input_113_beta_0_to_fp16, epsilon = input_113_epsilon_0_to_fp16, gamma = input_113_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_37_cast_fp16)[name = tensor<string, []>("input_113_cast_fp16")];
-            tensor<string, []> input_115_pad_type_0 = const()[name = tensor<string, []>("input_115_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_115_strides_0 = const()[name = tensor<string, []>("input_115_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_115_pad_0 = const()[name = tensor<string, []>("input_115_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_115_dilations_0 = const()[name = tensor<string, []>("input_115_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_115_groups_0 = const()[name = tensor<string, []>("input_115_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_1207_to_fp16 = const()[name = tensor<string, []>("op_1207_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_37_cast_fp16 = layer_norm(axes = out_37_axes_0, epsilon = var_1207_to_fp16, x = inputs_37_cast_fp16)[name = tensor<string, []>("out_37_cast_fp16")];
+            tensor<fp16, [1024]> input_115_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_115_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194177984)))];
+            tensor<fp16, [1024]> input_115_beta_0_to_fp16 = const()[name = tensor<string, []>("input_115_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194180096)))];
+            tensor<fp16, []> input_115_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_115_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_115_cast_fp16 = batch_norm(beta = input_115_beta_0_to_fp16, epsilon = input_115_epsilon_0_to_fp16, gamma = input_115_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_37_cast_fp16)[name = tensor<string, []>("input_115_cast_fp16")];
+            tensor<string, []> input_117_pad_type_0 = const()[name = tensor<string, []>("input_117_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_117_strides_0 = const()[name = tensor<string, []>("input_117_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_117_pad_0 = const()[name = tensor<string, []>("input_117_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_117_dilations_0 = const()[name = tensor<string, []>("input_117_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_117_groups_0 = const()[name = tensor<string, []>("input_117_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_3_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_feed_forward2_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194182208)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_115_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_115_dilations_0, groups = input_115_groups_0, pad = input_115_pad_0, pad_type = input_115_pad_type_0, strides = input_115_strides_0, weight = layers_3_feed_forward2_fc1_weight_to_fp16, x = input_113_cast_fp16)[name = tensor<string, []>("input_115_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_117_cast_fp16 = silu(x = input_115_cast_fp16)[name = tensor<string, []>("input_117_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_117_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_117_dilations_0, groups = input_117_groups_0, pad = input_117_pad_0, pad_type = input_117_pad_type_0, strides = input_117_strides_0, weight = layers_3_feed_forward2_fc1_weight_to_fp16, x = input_115_cast_fp16)[name = tensor<string, []>("input_117_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_119_cast_fp16 = silu(x = input_117_cast_fp16)[name = tensor<string, []>("input_119_cast_fp16")];
             tensor<string, []> x_25_pad_type_0 = const()[name = tensor<string, []>("x_25_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_25_strides_0 = const()[name = tensor<string, []>("x_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_25_pad_0 = const()[name = tensor<string, []>("x_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_25_dilations_0 = const()[name = tensor<string, []>("x_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_25_groups_0 = const()[name = tensor<string, []>("x_25_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_1237_weight_0_to_fp16 = const()[name = tensor<string, []>("op_1237_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202570880)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_1237_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_25_dilations_0, groups = x_25_groups_0, pad = x_25_pad_0, pad_type = x_25_pad_type_0, strides = x_25_strides_0, weight = var_1237_weight_0_to_fp16, x = input_117_cast_fp16)[name = tensor<string, []>("op_1237_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_39_cast_fp16 = add(x = inputs_37_cast_fp16, y = var_1237_cast_fp16)[name = tensor<string, []>("inputs_39_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_1235_weight_0_to_fp16 = const()[name = tensor<string, []>("op_1235_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202570880)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_1235_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_25_dilations_0, groups = x_25_groups_0, pad = x_25_pad_0, pad_type = x_25_pad_type_0, strides = x_25_strides_0, weight = var_1235_weight_0_to_fp16, x = input_119_cast_fp16)[name = tensor<string, []>("op_1235_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_39_cast_fp16 = add(x = inputs_37_cast_fp16, y = var_1235_cast_fp16)[name = tensor<string, []>("inputs_39_cast_fp16")];
             tensor<int32, [1]> out_39_axes_0 = const()[name = tensor<string, []>("out_39_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_1247_to_fp16 = const()[name = tensor<string, []>("op_1247_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_39_cast_fp16 = layer_norm(axes = out_39_axes_0, epsilon = var_1247_to_fp16, x = inputs_39_cast_fp16)[name = tensor<string, []>("out_39_cast_fp16")];
+            tensor<fp16, []> var_1245_to_fp16 = const()[name = tensor<string, []>("op_1245_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_39_cast_fp16 = layer_norm(axes = out_39_axes_0, epsilon = var_1245_to_fp16, x = inputs_39_cast_fp16)[name = tensor<string, []>("out_39_cast_fp16")];
             tensor<fp16, [1024]> inputs_41_gamma_0_to_fp16 = const()[name = tensor<string, []>("inputs_41_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(210959552)))];
             tensor<fp16, [1024]> inputs_41_beta_0_to_fp16 = const()[name = tensor<string, []>("inputs_41_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(210961664)))];
             tensor<fp16, []> inputs_41_epsilon_0_to_fp16 = const()[name = tensor<string, []>("inputs_41_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_41_cast_fp16 = batch_norm(beta = inputs_41_beta_0_to_fp16, epsilon = inputs_41_epsilon_0_to_fp16, gamma = inputs_41_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_39_cast_fp16)[name = tensor<string, []>("inputs_41_cast_fp16")];
-            tensor<int32, []> var_1261 = const()[name = tensor<string, []>("op_1261"), val = tensor<int32, []>(3)];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_41_cast_fp16 = batch_norm(beta = inputs_41_beta_0_to_fp16, epsilon = inputs_41_epsilon_0_to_fp16, gamma = inputs_41_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_39_cast_fp16)[name = tensor<string, []>("inputs_41_cast_fp16")];
+            tensor<int32, []> var_1259 = const()[name = tensor<string, []>("op_1259"), val = tensor<int32, []>(3)];
             tensor<int32, [1]> out_41_axes_0 = const()[name = tensor<string, []>("out_41_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_1292_to_fp16 = const()[name = tensor<string, []>("op_1292_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_41_cast_fp16 = layer_norm(axes = out_41_axes_0, epsilon = var_1292_to_fp16, x = inputs_41_cast_fp16)[name = tensor<string, []>("out_41_cast_fp16")];
-            tensor<fp16, [1024]> input_119_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_119_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(210963776)))];
-            tensor<fp16, [1024]> input_119_beta_0_to_fp16 = const()[name = tensor<string, []>("input_119_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(210965888)))];
-            tensor<fp16, []> input_119_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_119_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_119_cast_fp16 = batch_norm(beta = input_119_beta_0_to_fp16, epsilon = input_119_epsilon_0_to_fp16, gamma = input_119_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_41_cast_fp16)[name = tensor<string, []>("input_119_cast_fp16")];
-            tensor<string, []> input_121_pad_type_0 = const()[name = tensor<string, []>("input_121_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_121_strides_0 = const()[name = tensor<string, []>("input_121_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_121_pad_0 = const()[name = tensor<string, []>("input_121_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_121_dilations_0 = const()[name = tensor<string, []>("input_121_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_121_groups_0 = const()[name = tensor<string, []>("input_121_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_1290_to_fp16 = const()[name = tensor<string, []>("op_1290_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_41_cast_fp16 = layer_norm(axes = out_41_axes_0, epsilon = var_1290_to_fp16, x = inputs_41_cast_fp16)[name = tensor<string, []>("out_41_cast_fp16")];
+            tensor<fp16, [1024]> input_121_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_121_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(210963776)))];
+            tensor<fp16, [1024]> input_121_beta_0_to_fp16 = const()[name = tensor<string, []>("input_121_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(210965888)))];
+            tensor<fp16, []> input_121_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_121_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_121_cast_fp16 = batch_norm(beta = input_121_beta_0_to_fp16, epsilon = input_121_epsilon_0_to_fp16, gamma = input_121_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_41_cast_fp16)[name = tensor<string, []>("input_121_cast_fp16")];
+            tensor<string, []> input_123_pad_type_0 = const()[name = tensor<string, []>("input_123_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_123_strides_0 = const()[name = tensor<string, []>("input_123_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_123_pad_0 = const()[name = tensor<string, []>("input_123_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_123_dilations_0 = const()[name = tensor<string, []>("input_123_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_123_groups_0 = const()[name = tensor<string, []>("input_123_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_4_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_feed_forward1_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(210968000)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_121_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_121_dilations_0, groups = input_121_groups_0, pad = input_121_pad_0, pad_type = input_121_pad_type_0, strides = input_121_strides_0, weight = layers_4_feed_forward1_fc1_weight_to_fp16, x = input_119_cast_fp16)[name = tensor<string, []>("input_121_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_123_cast_fp16 = silu(x = input_121_cast_fp16)[name = tensor<string, []>("input_123_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_123_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_123_dilations_0, groups = input_123_groups_0, pad = input_123_pad_0, pad_type = input_123_pad_type_0, strides = input_123_strides_0, weight = layers_4_feed_forward1_fc1_weight_to_fp16, x = input_121_cast_fp16)[name = tensor<string, []>("input_123_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_125_cast_fp16 = silu(x = input_123_cast_fp16)[name = tensor<string, []>("input_125_cast_fp16")];
             tensor<string, []> x_27_pad_type_0 = const()[name = tensor<string, []>("x_27_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_27_strides_0 = const()[name = tensor<string, []>("x_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_27_pad_0 = const()[name = tensor<string, []>("x_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_27_dilations_0 = const()[name = tensor<string, []>("x_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_27_groups_0 = const()[name = tensor<string, []>("x_27_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_1320_weight_0_to_fp16 = const()[name = tensor<string, []>("op_1320_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(219356672)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_1320_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_27_dilations_0, groups = x_27_groups_0, pad = x_27_pad_0, pad_type = x_27_pad_type_0, strides = x_27_strides_0, weight = var_1320_weight_0_to_fp16, x = input_123_cast_fp16)[name = tensor<string, []>("op_1320_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_43_cast_fp16 = add(x = inputs_41_cast_fp16, y = var_1320_cast_fp16)[name = tensor<string, []>("inputs_43_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_1318_weight_0_to_fp16 = const()[name = tensor<string, []>("op_1318_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(219356672)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_1318_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_27_dilations_0, groups = x_27_groups_0, pad = x_27_pad_0, pad_type = x_27_pad_type_0, strides = x_27_strides_0, weight = var_1318_weight_0_to_fp16, x = input_125_cast_fp16)[name = tensor<string, []>("op_1318_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_43_cast_fp16 = add(x = inputs_41_cast_fp16, y = var_1318_cast_fp16)[name = tensor<string, []>("inputs_43_cast_fp16")];
             tensor<int32, [1]> out_43_axes_0 = const()[name = tensor<string, []>("out_43_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_1330_to_fp16 = const()[name = tensor<string, []>("op_1330_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_43_cast_fp16 = layer_norm(axes = out_43_axes_0, epsilon = var_1330_to_fp16, x = inputs_43_cast_fp16)[name = tensor<string, []>("out_43_cast_fp16")];
+            tensor<fp16, []> var_1328_to_fp16 = const()[name = tensor<string, []>("op_1328_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_43_cast_fp16 = layer_norm(axes = out_43_axes_0, epsilon = var_1328_to_fp16, x = inputs_43_cast_fp16)[name = tensor<string, []>("out_43_cast_fp16")];
             tensor<fp16, [1024]> obj_19_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_19_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(227745344)))];
             tensor<fp16, [1024]> obj_19_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_19_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(227747456)))];
             tensor<fp16, []> obj_19_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_19_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> obj_19_cast_fp16 = batch_norm(beta = obj_19_beta_0_to_fp16, epsilon = obj_19_epsilon_0_to_fp16, gamma = obj_19_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_43_cast_fp16)[name = tensor<string, []>("obj_19_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_19_cast_fp16 = batch_norm(beta = obj_19_beta_0_to_fp16, epsilon = obj_19_epsilon_0_to_fp16, gamma = obj_19_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_43_cast_fp16)[name = tensor<string, []>("obj_19_cast_fp16")];
             tensor<string, []> query_17_pad_type_0 = const()[name = tensor<string, []>("query_17_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> query_17_strides_0 = const()[name = tensor<string, []>("query_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> query_17_pad_0 = const()[name = tensor<string, []>("query_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> query_17_dilations_0 = const()[name = tensor<string, []>("query_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> query_17_groups_0 = const()[name = tensor<string, []>("query_17_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_4_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(227749568)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_17_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = query_17_dilations_0, groups = query_17_groups_0, pad = query_17_pad_0, pad_type = query_17_pad_type_0, strides = query_17_strides_0, weight = layers_4_self_attn_q_proj_weight_to_fp16, x = obj_19_cast_fp16)[name = tensor<string, []>("query_17_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> query_17_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_17_dilations_0, groups = query_17_groups_0, pad = query_17_pad_0, pad_type = query_17_pad_type_0, strides = query_17_strides_0, weight = layers_4_self_attn_q_proj_weight_to_fp16, x = obj_19_cast_fp16)[name = tensor<string, []>("query_17_cast_fp16")];
             tensor<string, []> key_9_pad_type_0 = const()[name = tensor<string, []>("key_9_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> key_9_strides_0 = const()[name = tensor<string, []>("key_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> key_9_pad_0 = const()[name = tensor<string, []>("key_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
@@ -832,173 +834,173 @@ program(1.0)
             tensor<int32, [2]> value_9_dilations_0 = const()[name = tensor<string, []>("value_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> value_9_groups_0 = const()[name = tensor<string, []>("value_9_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_4_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231944000)))];
-            tensor<fp16, [1, 1024, 1, 188]> value_9_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = value_9_dilations_0, groups = value_9_groups_0, pad = value_9_pad_0, pad_type = value_9_pad_type_0, strides = value_9_strides_0, weight = layers_4_self_attn_v_proj_weight_to_fp16, x = obj_19_cast_fp16)[name = tensor<string, []>("value_9_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_1368_to_fp16 = const()[name = tensor<string, []>("op_1368_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(234041216)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_19_cast_fp16 = add(x = query_17_cast_fp16, y = var_1368_to_fp16)[name = tensor<string, []>("query_19_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_1371_to_fp16 = const()[name = tensor<string, []>("op_1371_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(234043328)))];
-            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_9_cast_fp16 = add(x = query_17_cast_fp16, y = var_1371_to_fp16)[name = tensor<string, []>("q_with_bias_v_9_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> value_9_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_9_dilations_0, groups = value_9_groups_0, pad = value_9_pad_0, pad_type = value_9_pad_type_0, strides = value_9_strides_0, weight = layers_4_self_attn_v_proj_weight_to_fp16, x = obj_19_cast_fp16)[name = tensor<string, []>("value_9_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_1366_to_fp16 = const()[name = tensor<string, []>("op_1366_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(234041216)))];
+            tensor<fp16, [1, 1024, 1, 188]> query_19_cast_fp16 = add(x = query_17_cast_fp16, y = var_1366_to_fp16)[name = tensor<string, []>("query_19_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_1369_to_fp16 = const()[name = tensor<string, []>("op_1369_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(234043328)))];
+            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_9_cast_fp16 = add(x = query_17_cast_fp16, y = var_1369_to_fp16)[name = tensor<string, []>("q_with_bias_v_9_cast_fp16")];
             tensor<string, []> p_9_pad_type_0 = const()[name = tensor<string, []>("p_9_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> p_9_strides_0 = const()[name = tensor<string, []>("p_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> p_9_pad_0 = const()[name = tensor<string, []>("p_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> p_9_dilations_0 = const()[name = tensor<string, []>("p_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> p_9_groups_0 = const()[name = tensor<string, []>("p_9_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_4_self_attn_linear_pos_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_linear_pos_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(234045440)))];
-            tensor<fp16, [1, 1024, 1, 375]> p_9_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = p_9_dilations_0, groups = p_9_groups_0, pad = p_9_pad_0, pad_type = p_9_pad_type_0, strides = p_9_strides_0, weight = layers_4_self_attn_linear_pos_weight_to_fp16, x = pos_enc_to_fp16)[name = tensor<string, []>("p_9_cast_fp16")];
-            tensor<int32, [4]> var_1382 = const()[name = tensor<string, []>("op_1382"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_1383_cast_fp16 = reshape(shape = var_1382, x = q_with_bias_v_9_cast_fp16)[name = tensor<string, []>("op_1383_cast_fp16")];
-            tensor<int32, [4]> var_1384 = const()[name = tensor<string, []>("op_1384"), val = tensor<int32, [4]>([1, 8, 128, -1])];
-            tensor<fp16, [1, 8, 128, 375]> var_1385_cast_fp16 = reshape(shape = var_1384, x = p_9_cast_fp16)[name = tensor<string, []>("op_1385_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 375]> p_9_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_9_dilations_0, groups = p_9_groups_0, pad = p_9_pad_0, pad_type = p_9_pad_type_0, strides = p_9_strides_0, weight = layers_4_self_attn_linear_pos_weight_to_fp16, x = obj_3_cast_fp16)[name = tensor<string, []>("p_9_cast_fp16")];
+            tensor<int32, [4]> var_1380 = const()[name = tensor<string, []>("op_1380"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_1381_cast_fp16 = reshape(shape = var_1380, x = q_with_bias_v_9_cast_fp16)[name = tensor<string, []>("op_1381_cast_fp16")];
+            tensor<int32, [4]> var_1382 = const()[name = tensor<string, []>("op_1382"), val = tensor<int32, [4]>([1, 8, 128, -1])];
+            tensor<fp16, [1, 8, 128, 375]> var_1383_cast_fp16 = reshape(shape = var_1382, x = p_9_cast_fp16)[name = tensor<string, []>("op_1383_cast_fp16")];
             tensor<bool, []> matrix_bd_33_transpose_x_0 = const()[name = tensor<string, []>("matrix_bd_33_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> matrix_bd_33_transpose_y_0 = const()[name = tensor<string, []>("matrix_bd_33_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_33_cast_fp16 = matmul(transpose_x = matrix_bd_33_transpose_x_0, transpose_y = matrix_bd_33_transpose_y_0, x = var_1383_cast_fp16, y = var_1385_cast_fp16)[name = tensor<string, []>("matrix_bd_33_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_33_cast_fp16 = matmul(transpose_x = matrix_bd_33_transpose_x_0, transpose_y = matrix_bd_33_transpose_y_0, x = var_1381_cast_fp16, y = var_1383_cast_fp16)[name = tensor<string, []>("matrix_bd_33_cast_fp16")];
             tensor<int32, [8]> matrix_bd_35_pad_0 = const()[name = tensor<string, []>("matrix_bd_35_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
             tensor<string, []> matrix_bd_35_mode_0 = const()[name = tensor<string, []>("matrix_bd_35_mode_0"), val = tensor<string, []>("constant")];
             tensor<fp16, []> const_54_to_fp16 = const()[name = tensor<string, []>("const_54_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
             tensor<fp16, [1, 8, 188, 376]> matrix_bd_35_cast_fp16 = pad(constant_val = const_54_to_fp16, mode = matrix_bd_35_mode_0, pad = matrix_bd_35_pad_0, x = matrix_bd_33_cast_fp16)[name = tensor<string, []>("matrix_bd_35_cast_fp16")];
-            tensor<int32, [4]> var_1394 = const()[name = tensor<string, []>("op_1394"), val = tensor<int32, [4]>([1, 8, -1, 188])];
-            tensor<fp16, [1, 8, 376, 188]> matrix_bd_37_cast_fp16 = reshape(shape = var_1394, x = matrix_bd_35_cast_fp16)[name = tensor<string, []>("matrix_bd_37_cast_fp16")];
-            tensor<int32, [4]> var_1398_begin_0 = const()[name = tensor<string, []>("op_1398_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
-            tensor<int32, [4]> var_1398_end_0 = const()[name = tensor<string, []>("op_1398_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
-            tensor<bool, [4]> var_1398_end_mask_0 = const()[name = tensor<string, []>("op_1398_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
-            tensor<fp16, [1, 8, 375, 188]> var_1398_cast_fp16 = slice_by_index(begin = var_1398_begin_0, end = var_1398_end_0, end_mask = var_1398_end_mask_0, x = matrix_bd_37_cast_fp16)[name = tensor<string, []>("op_1398_cast_fp16")];
-            tensor<int32, [4]> var_1399 = const()[name = tensor<string, []>("op_1399"), val = tensor<int32, [4]>([1, 8, 188, 375])];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_39_cast_fp16 = reshape(shape = var_1399, x = var_1398_cast_fp16)[name = tensor<string, []>("matrix_bd_39_cast_fp16")];
-            tensor<int32, [4]> var_1404_begin_0 = const()[name = tensor<string, []>("op_1404_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [4]> var_1404_end_0 = const()[name = tensor<string, []>("op_1404_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
-            tensor<bool, [4]> var_1404_end_mask_0 = const()[name = tensor<string, []>("op_1404_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
-            tensor<fp16, [1, 8, 188, 188]> var_1404_cast_fp16 = slice_by_index(begin = var_1404_begin_0, end = var_1404_end_0, end_mask = var_1404_end_mask_0, x = matrix_bd_39_cast_fp16)[name = tensor<string, []>("op_1404_cast_fp16")];
-            tensor<fp16, []> var_1405_to_fp16 = const()[name = tensor<string, []>("op_1405_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 188, 188]> qk_mask_9_cast_fp16 = mul(x = var_1404_cast_fp16, y = var_1405_to_fp16)[name = tensor<string, []>("qk_mask_9_cast_fp16")];
-            tensor<int32, [4]> var_1409 = const()[name = tensor<string, []>("op_1409"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> mh_q_9_cast_fp16 = reshape(shape = var_1409, x = query_19_cast_fp16)[name = tensor<string, []>("mh_q_9_cast_fp16")];
-            tensor<fp16, []> var_1411_to_fp16 = const()[name = tensor<string, []>("op_1411_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 128, 188]> var_1412_cast_fp16 = mul(x = mh_q_9_cast_fp16, y = var_1411_to_fp16)[name = tensor<string, []>("op_1412_cast_fp16")];
-            tensor<int32, [4]> var_1415 = const()[name = tensor<string, []>("op_1415"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_1416_cast_fp16 = reshape(shape = var_1415, x = key_9_cast_fp16)[name = tensor<string, []>("op_1416_cast_fp16")];
+            tensor<int32, [4]> var_1392 = const()[name = tensor<string, []>("op_1392"), val = tensor<int32, [4]>([1, 8, -1, 188])];
+            tensor<fp16, [1, 8, 376, 188]> matrix_bd_37_cast_fp16 = reshape(shape = var_1392, x = matrix_bd_35_cast_fp16)[name = tensor<string, []>("matrix_bd_37_cast_fp16")];
+            tensor<int32, [4]> var_1396_begin_0 = const()[name = tensor<string, []>("op_1396_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1396_end_0 = const()[name = tensor<string, []>("op_1396_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
+            tensor<bool, [4]> var_1396_end_mask_0 = const()[name = tensor<string, []>("op_1396_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 375, 188]> var_1396_cast_fp16 = slice_by_index(begin = var_1396_begin_0, end = var_1396_end_0, end_mask = var_1396_end_mask_0, x = matrix_bd_37_cast_fp16)[name = tensor<string, []>("op_1396_cast_fp16")];
+            tensor<int32, [4]> var_1397 = const()[name = tensor<string, []>("op_1397"), val = tensor<int32, [4]>([1, 8, 188, 375])];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_39_cast_fp16 = reshape(shape = var_1397, x = var_1396_cast_fp16)[name = tensor<string, []>("matrix_bd_39_cast_fp16")];
+            tensor<int32, [4]> var_1402_begin_0 = const()[name = tensor<string, []>("op_1402_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_1402_end_0 = const()[name = tensor<string, []>("op_1402_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
+            tensor<bool, [4]> var_1402_end_mask_0 = const()[name = tensor<string, []>("op_1402_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 188, 188]> var_1402_cast_fp16 = slice_by_index(begin = var_1402_begin_0, end = var_1402_end_0, end_mask = var_1402_end_mask_0, x = matrix_bd_39_cast_fp16)[name = tensor<string, []>("op_1402_cast_fp16")];
+            tensor<fp16, []> var_1403_to_fp16 = const()[name = tensor<string, []>("op_1403_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 188, 188]> qk_mask_9_cast_fp16 = mul(x = var_1402_cast_fp16, y = var_1403_to_fp16)[name = tensor<string, []>("qk_mask_9_cast_fp16")];
+            tensor<int32, [4]> var_1407 = const()[name = tensor<string, []>("op_1407"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> mh_q_9_cast_fp16 = reshape(shape = var_1407, x = query_19_cast_fp16)[name = tensor<string, []>("mh_q_9_cast_fp16")];
+            tensor<fp16, []> var_1409_to_fp16 = const()[name = tensor<string, []>("op_1409_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 128, 188]> var_1410_cast_fp16 = mul(x = mh_q_9_cast_fp16, y = var_1409_to_fp16)[name = tensor<string, []>("op_1410_cast_fp16")];
+            tensor<int32, [4]> var_1413 = const()[name = tensor<string, []>("op_1413"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_1414_cast_fp16 = reshape(shape = var_1413, x = key_9_cast_fp16)[name = tensor<string, []>("op_1414_cast_fp16")];
             tensor<bool, []> mh_w_17_transpose_x_0 = const()[name = tensor<string, []>("mh_w_17_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> mh_w_17_transpose_y_0 = const()[name = tensor<string, []>("mh_w_17_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 188]> mh_w_17_cast_fp16 = matmul(transpose_x = mh_w_17_transpose_x_0, transpose_y = mh_w_17_transpose_y_0, x = var_1412_cast_fp16, y = var_1416_cast_fp16)[name = tensor<string, []>("mh_w_17_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> mh_w_17_cast_fp16 = matmul(transpose_x = mh_w_17_transpose_x_0, transpose_y = mh_w_17_transpose_y_0, x = var_1410_cast_fp16, y = var_1414_cast_fp16)[name = tensor<string, []>("mh_w_17_cast_fp16")];
             tensor<fp16, [1, 8, 188, 188]> mh_w_19_cast_fp16 = add(x = mh_w_17_cast_fp16, y = qk_mask_9_cast_fp16)[name = tensor<string, []>("mh_w_19_cast_fp16")];
-            tensor<fp16, [1, 8, 188, 188]> var_1420_cast_fp16 = softmax(axis = var_1261, x = mh_w_19_cast_fp16)[name = tensor<string, []>("op_1420_cast_fp16")];
-            tensor<int32, [4]> var_1421 = const()[name = tensor<string, []>("op_1421"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_1422_cast_fp16 = reshape(shape = var_1421, x = value_9_cast_fp16)[name = tensor<string, []>("op_1422_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> var_1418_cast_fp16 = softmax(axis = var_1259, x = mh_w_19_cast_fp16)[name = tensor<string, []>("op_1418_cast_fp16")];
+            tensor<int32, [4]> var_1419 = const()[name = tensor<string, []>("op_1419"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_1420_cast_fp16 = reshape(shape = var_1419, x = value_9_cast_fp16)[name = tensor<string, []>("op_1420_cast_fp16")];
             tensor<bool, []> attn_9_transpose_x_0 = const()[name = tensor<string, []>("attn_9_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> attn_9_transpose_y_0 = const()[name = tensor<string, []>("attn_9_transpose_y_0"), val = tensor<bool, []>(true)];
-            tensor<fp16, [1, 8, 128, 188]> attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = var_1422_cast_fp16, y = var_1420_cast_fp16)[name = tensor<string, []>("attn_9_cast_fp16")];
-            tensor<int32, [4]> var_1425 = const()[name = tensor<string, []>("op_1425"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
-            tensor<fp16, [1, 1024, 1, 188]> input_125_cast_fp16 = reshape(shape = var_1425, x = attn_9_cast_fp16)[name = tensor<string, []>("input_125_cast_fp16")];
+            tensor<fp16, [1, 8, 128, 188]> attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = var_1420_cast_fp16, y = var_1418_cast_fp16)[name = tensor<string, []>("attn_9_cast_fp16")];
+            tensor<int32, [4]> var_1423 = const()[name = tensor<string, []>("op_1423"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
+            tensor<fp16, [1, 1024, 1, 188]> input_127_cast_fp16 = reshape(shape = var_1423, x = attn_9_cast_fp16)[name = tensor<string, []>("input_127_cast_fp16")];
             tensor<string, []> obj_21_pad_type_0 = const()[name = tensor<string, []>("obj_21_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> obj_21_strides_0 = const()[name = tensor<string, []>("obj_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> obj_21_pad_0 = const()[name = tensor<string, []>("obj_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> obj_21_dilations_0 = const()[name = tensor<string, []>("obj_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> obj_21_groups_0 = const()[name = tensor<string, []>("obj_21_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_4_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(236142656)))];
-            tensor<fp16, [1, 1024, 1, 188]> obj_21_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = obj_21_dilations_0, groups = obj_21_groups_0, pad = obj_21_pad_0, pad_type = obj_21_pad_type_0, strides = obj_21_strides_0, weight = layers_4_self_attn_o_proj_weight_to_fp16, x = input_125_cast_fp16)[name = tensor<string, []>("obj_21_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_21_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_21_dilations_0, groups = obj_21_groups_0, pad = obj_21_pad_0, pad_type = obj_21_pad_type_0, strides = obj_21_strides_0, weight = layers_4_self_attn_o_proj_weight_to_fp16, x = input_127_cast_fp16)[name = tensor<string, []>("obj_21_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_45_cast_fp16 = add(x = inputs_43_cast_fp16, y = obj_21_cast_fp16)[name = tensor<string, []>("inputs_45_cast_fp16")];
             tensor<int32, [1]> out_45_axes_0 = const()[name = tensor<string, []>("out_45_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_1443_to_fp16 = const()[name = tensor<string, []>("op_1443_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_45_cast_fp16 = layer_norm(axes = out_45_axes_0, epsilon = var_1443_to_fp16, x = inputs_45_cast_fp16)[name = tensor<string, []>("out_45_cast_fp16")];
-            tensor<fp16, [1024]> input_127_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_127_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(238239872)))];
-            tensor<fp16, [1024]> input_127_beta_0_to_fp16 = const()[name = tensor<string, []>("input_127_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(238241984)))];
-            tensor<fp16, []> input_127_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_127_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_127_cast_fp16 = batch_norm(beta = input_127_beta_0_to_fp16, epsilon = input_127_epsilon_0_to_fp16, gamma = input_127_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_45_cast_fp16)[name = tensor<string, []>("input_127_cast_fp16")];
-            tensor<string, []> input_129_pad_type_0 = const()[name = tensor<string, []>("input_129_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_129_strides_0 = const()[name = tensor<string, []>("input_129_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_129_pad_0 = const()[name = tensor<string, []>("input_129_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_129_dilations_0 = const()[name = tensor<string, []>("input_129_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_129_groups_0 = const()[name = tensor<string, []>("input_129_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_1441_to_fp16 = const()[name = tensor<string, []>("op_1441_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_45_cast_fp16 = layer_norm(axes = out_45_axes_0, epsilon = var_1441_to_fp16, x = inputs_45_cast_fp16)[name = tensor<string, []>("out_45_cast_fp16")];
+            tensor<fp16, [1024]> input_129_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_129_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(238239872)))];
+            tensor<fp16, [1024]> input_129_beta_0_to_fp16 = const()[name = tensor<string, []>("input_129_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(238241984)))];
+            tensor<fp16, []> input_129_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_129_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_129_cast_fp16 = batch_norm(beta = input_129_beta_0_to_fp16, epsilon = input_129_epsilon_0_to_fp16, gamma = input_129_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_45_cast_fp16)[name = tensor<string, []>("input_129_cast_fp16")];
+            tensor<string, []> input_131_pad_type_0 = const()[name = tensor<string, []>("input_131_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_131_strides_0 = const()[name = tensor<string, []>("input_131_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_131_pad_0 = const()[name = tensor<string, []>("input_131_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_131_dilations_0 = const()[name = tensor<string, []>("input_131_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_131_groups_0 = const()[name = tensor<string, []>("input_131_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [2048, 1024, 1, 1]> layers_4_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(238244096)))];
-            tensor<fp16, [1, 2048, 1, 188]> input_129_cast_fp16 = conv(dilations = input_129_dilations_0, groups = input_129_groups_0, pad = input_129_pad_0, pad_type = input_129_pad_type_0, strides = input_129_strides_0, weight = layers_4_conv_pointwise_conv1_weight_to_fp16, x = input_127_cast_fp16)[name = tensor<string, []>("input_129_cast_fp16")];
-            tensor<int32, []> input_131_split_num_splits_0 = const()[name = tensor<string, []>("input_131_split_num_splits_0"), val = tensor<int32, []>(2)];
-            tensor<int32, []> input_131_split_axis_0 = const()[name = tensor<string, []>("input_131_split_axis_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1, 1024, 1, 188]> input_131_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_131_split_cast_fp16_1 = split(axis = input_131_split_axis_0, num_splits = input_131_split_num_splits_0, x = input_129_cast_fp16)[name = tensor<string, []>("input_131_split_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_131_split_1_sigmoid_cast_fp16 = sigmoid(x = input_131_split_cast_fp16_1)[name = tensor<string, []>("input_131_split_1_sigmoid_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_131_cast_fp16 = mul(x = input_131_split_cast_fp16_0, y = input_131_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_131_cast_fp16")];
-            tensor<string, []> input_133_pad_type_0 = const()[name = tensor<string, []>("input_133_pad_type_0"), val = tensor<string, []>("custom")];
-            tensor<int32, [4]> input_133_pad_0 = const()[name = tensor<string, []>("input_133_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
-            tensor<int32, []> input_133_groups_0 = const()[name = tensor<string, []>("input_133_groups_0"), val = tensor<int32, []>(1024)];
-            tensor<int32, [2]> input_133_strides_0 = const()[name = tensor<string, []>("input_133_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> input_133_dilations_0 = const()[name = tensor<string, []>("input_133_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [1, 2048, 1, 188]> input_131_cast_fp16 = conv(dilations = input_131_dilations_0, groups = input_131_groups_0, pad = input_131_pad_0, pad_type = input_131_pad_type_0, strides = input_131_strides_0, weight = layers_4_conv_pointwise_conv1_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("input_131_cast_fp16")];
+            tensor<int32, []> input_133_split_num_splits_0 = const()[name = tensor<string, []>("input_133_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> input_133_split_axis_0 = const()[name = tensor<string, []>("input_133_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 1, 188]> input_133_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_133_split_cast_fp16_1 = split(axis = input_133_split_axis_0, num_splits = input_133_split_num_splits_0, x = input_131_cast_fp16)[name = tensor<string, []>("input_133_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_133_split_1_sigmoid_cast_fp16 = sigmoid(x = input_133_split_cast_fp16_1)[name = tensor<string, []>("input_133_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_133_cast_fp16 = mul(x = input_133_split_cast_fp16_0, y = input_133_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_133_cast_fp16")];
+            tensor<string, []> input_135_pad_type_0 = const()[name = tensor<string, []>("input_135_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_135_pad_0 = const()[name = tensor<string, []>("input_135_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
+            tensor<int32, []> input_135_groups_0 = const()[name = tensor<string, []>("input_135_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [2]> input_135_strides_0 = const()[name = tensor<string, []>("input_135_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> input_135_dilations_0 = const()[name = tensor<string, []>("input_135_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<fp16, [1024, 1, 1, 9]> const_276_to_fp16 = const()[name = tensor<string, []>("const_276_to_fp16"), val = tensor<fp16, [1024, 1, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(242438464)))];
             tensor<fp16, [1024]> const_277_to_fp16 = const()[name = tensor<string, []>("const_277_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(242456960)))];
-            tensor<fp16, [1, 1024, 1, 188]> input_135_cast_fp16 = conv(bias = const_277_to_fp16, dilations = input_133_dilations_0, groups = input_133_groups_0, pad = input_133_pad_0, pad_type = input_133_pad_type_0, strides = input_133_strides_0, weight = const_276_to_fp16, x = input_131_cast_fp16)[name = tensor<string, []>("input_135_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_137_cast_fp16 = silu(x = input_135_cast_fp16)[name = tensor<string, []>("input_137_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_137_cast_fp16 = conv(bias = const_277_to_fp16, dilations = input_135_dilations_0, groups = input_135_groups_0, pad = input_135_pad_0, pad_type = input_135_pad_type_0, strides = input_135_strides_0, weight = const_276_to_fp16, x = input_133_cast_fp16)[name = tensor<string, []>("input_137_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_139_cast_fp16 = silu(x = input_137_cast_fp16)[name = tensor<string, []>("input_139_cast_fp16")];
             tensor<string, []> x_29_pad_type_0 = const()[name = tensor<string, []>("x_29_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_29_strides_0 = const()[name = tensor<string, []>("x_29_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_29_pad_0 = const()[name = tensor<string, []>("x_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_29_dilations_0 = const()[name = tensor<string, []>("x_29_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_29_groups_0 = const()[name = tensor<string, []>("x_29_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_4_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(242459072)))];
-            tensor<fp16, [1, 1024, 1, 188]> x_29_cast_fp16 = conv(dilations = x_29_dilations_0, groups = x_29_groups_0, pad = x_29_pad_0, pad_type = x_29_pad_type_0, strides = x_29_strides_0, weight = layers_4_conv_pointwise_conv2_weight_to_fp16, x = input_137_cast_fp16)[name = tensor<string, []>("x_29_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> x_29_cast_fp16 = conv(dilations = x_29_dilations_0, groups = x_29_groups_0, pad = x_29_pad_0, pad_type = x_29_pad_type_0, strides = x_29_strides_0, weight = layers_4_conv_pointwise_conv2_weight_to_fp16, x = input_139_cast_fp16)[name = tensor<string, []>("x_29_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_47_cast_fp16 = add(x = inputs_45_cast_fp16, y = x_29_cast_fp16)[name = tensor<string, []>("inputs_47_cast_fp16")];
             tensor<int32, [1]> out_47_axes_0 = const()[name = tensor<string, []>("out_47_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_1491_to_fp16 = const()[name = tensor<string, []>("op_1491_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_47_cast_fp16 = layer_norm(axes = out_47_axes_0, epsilon = var_1491_to_fp16, x = inputs_47_cast_fp16)[name = tensor<string, []>("out_47_cast_fp16")];
-            tensor<fp16, [1024]> input_139_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_139_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(244556288)))];
-            tensor<fp16, [1024]> input_139_beta_0_to_fp16 = const()[name = tensor<string, []>("input_139_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(244558400)))];
-            tensor<fp16, []> input_139_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_139_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_139_cast_fp16 = batch_norm(beta = input_139_beta_0_to_fp16, epsilon = input_139_epsilon_0_to_fp16, gamma = input_139_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_47_cast_fp16)[name = tensor<string, []>("input_139_cast_fp16")];
-            tensor<string, []> input_141_pad_type_0 = const()[name = tensor<string, []>("input_141_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_141_strides_0 = const()[name = tensor<string, []>("input_141_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_141_pad_0 = const()[name = tensor<string, []>("input_141_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_141_dilations_0 = const()[name = tensor<string, []>("input_141_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_141_groups_0 = const()[name = tensor<string, []>("input_141_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_1489_to_fp16 = const()[name = tensor<string, []>("op_1489_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_47_cast_fp16 = layer_norm(axes = out_47_axes_0, epsilon = var_1489_to_fp16, x = inputs_47_cast_fp16)[name = tensor<string, []>("out_47_cast_fp16")];
+            tensor<fp16, [1024]> input_141_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_141_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(244556288)))];
+            tensor<fp16, [1024]> input_141_beta_0_to_fp16 = const()[name = tensor<string, []>("input_141_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(244558400)))];
+            tensor<fp16, []> input_141_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_141_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_141_cast_fp16 = batch_norm(beta = input_141_beta_0_to_fp16, epsilon = input_141_epsilon_0_to_fp16, gamma = input_141_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_47_cast_fp16)[name = tensor<string, []>("input_141_cast_fp16")];
+            tensor<string, []> input_143_pad_type_0 = const()[name = tensor<string, []>("input_143_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_143_strides_0 = const()[name = tensor<string, []>("input_143_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_143_pad_0 = const()[name = tensor<string, []>("input_143_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_143_dilations_0 = const()[name = tensor<string, []>("input_143_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_143_groups_0 = const()[name = tensor<string, []>("input_143_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_4_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_feed_forward2_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(244560512)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_141_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_141_dilations_0, groups = input_141_groups_0, pad = input_141_pad_0, pad_type = input_141_pad_type_0, strides = input_141_strides_0, weight = layers_4_feed_forward2_fc1_weight_to_fp16, x = input_139_cast_fp16)[name = tensor<string, []>("input_141_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_143_cast_fp16 = silu(x = input_141_cast_fp16)[name = tensor<string, []>("input_143_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_143_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_143_dilations_0, groups = input_143_groups_0, pad = input_143_pad_0, pad_type = input_143_pad_type_0, strides = input_143_strides_0, weight = layers_4_feed_forward2_fc1_weight_to_fp16, x = input_141_cast_fp16)[name = tensor<string, []>("input_143_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_145_cast_fp16 = silu(x = input_143_cast_fp16)[name = tensor<string, []>("input_145_cast_fp16")];
             tensor<string, []> x_31_pad_type_0 = const()[name = tensor<string, []>("x_31_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_31_strides_0 = const()[name = tensor<string, []>("x_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_31_pad_0 = const()[name = tensor<string, []>("x_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_31_dilations_0 = const()[name = tensor<string, []>("x_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_31_groups_0 = const()[name = tensor<string, []>("x_31_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_1519_weight_0_to_fp16 = const()[name = tensor<string, []>("op_1519_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(252949184)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_1519_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_31_dilations_0, groups = x_31_groups_0, pad = x_31_pad_0, pad_type = x_31_pad_type_0, strides = x_31_strides_0, weight = var_1519_weight_0_to_fp16, x = input_143_cast_fp16)[name = tensor<string, []>("op_1519_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_49_cast_fp16 = add(x = inputs_47_cast_fp16, y = var_1519_cast_fp16)[name = tensor<string, []>("inputs_49_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_1517_weight_0_to_fp16 = const()[name = tensor<string, []>("op_1517_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(252949184)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_1517_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_31_dilations_0, groups = x_31_groups_0, pad = x_31_pad_0, pad_type = x_31_pad_type_0, strides = x_31_strides_0, weight = var_1517_weight_0_to_fp16, x = input_145_cast_fp16)[name = tensor<string, []>("op_1517_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_49_cast_fp16 = add(x = inputs_47_cast_fp16, y = var_1517_cast_fp16)[name = tensor<string, []>("inputs_49_cast_fp16")];
             tensor<int32, [1]> out_49_axes_0 = const()[name = tensor<string, []>("out_49_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_1529_to_fp16 = const()[name = tensor<string, []>("op_1529_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_49_cast_fp16 = layer_norm(axes = out_49_axes_0, epsilon = var_1529_to_fp16, x = inputs_49_cast_fp16)[name = tensor<string, []>("out_49_cast_fp16")];
+            tensor<fp16, []> var_1527_to_fp16 = const()[name = tensor<string, []>("op_1527_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_49_cast_fp16 = layer_norm(axes = out_49_axes_0, epsilon = var_1527_to_fp16, x = inputs_49_cast_fp16)[name = tensor<string, []>("out_49_cast_fp16")];
             tensor<fp16, [1024]> inputs_51_gamma_0_to_fp16 = const()[name = tensor<string, []>("inputs_51_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(261337856)))];
             tensor<fp16, [1024]> inputs_51_beta_0_to_fp16 = const()[name = tensor<string, []>("inputs_51_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(261339968)))];
             tensor<fp16, []> inputs_51_epsilon_0_to_fp16 = const()[name = tensor<string, []>("inputs_51_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_51_cast_fp16 = batch_norm(beta = inputs_51_beta_0_to_fp16, epsilon = inputs_51_epsilon_0_to_fp16, gamma = inputs_51_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_49_cast_fp16)[name = tensor<string, []>("inputs_51_cast_fp16")];
-            tensor<int32, []> var_1543 = const()[name = tensor<string, []>("op_1543"), val = tensor<int32, []>(3)];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_51_cast_fp16 = batch_norm(beta = inputs_51_beta_0_to_fp16, epsilon = inputs_51_epsilon_0_to_fp16, gamma = inputs_51_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_49_cast_fp16)[name = tensor<string, []>("inputs_51_cast_fp16")];
+            tensor<int32, []> var_1541 = const()[name = tensor<string, []>("op_1541"), val = tensor<int32, []>(3)];
             tensor<int32, [1]> out_51_axes_0 = const()[name = tensor<string, []>("out_51_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_1574_to_fp16 = const()[name = tensor<string, []>("op_1574_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_51_cast_fp16 = layer_norm(axes = out_51_axes_0, epsilon = var_1574_to_fp16, x = inputs_51_cast_fp16)[name = tensor<string, []>("out_51_cast_fp16")];
-            tensor<fp16, [1024]> input_145_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_145_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(261342080)))];
-            tensor<fp16, [1024]> input_145_beta_0_to_fp16 = const()[name = tensor<string, []>("input_145_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(261344192)))];
-            tensor<fp16, []> input_145_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_145_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_145_cast_fp16 = batch_norm(beta = input_145_beta_0_to_fp16, epsilon = input_145_epsilon_0_to_fp16, gamma = input_145_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_51_cast_fp16)[name = tensor<string, []>("input_145_cast_fp16")];
-            tensor<string, []> input_147_pad_type_0 = const()[name = tensor<string, []>("input_147_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_147_strides_0 = const()[name = tensor<string, []>("input_147_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_147_pad_0 = const()[name = tensor<string, []>("input_147_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_147_dilations_0 = const()[name = tensor<string, []>("input_147_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_147_groups_0 = const()[name = tensor<string, []>("input_147_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_1572_to_fp16 = const()[name = tensor<string, []>("op_1572_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_51_cast_fp16 = layer_norm(axes = out_51_axes_0, epsilon = var_1572_to_fp16, x = inputs_51_cast_fp16)[name = tensor<string, []>("out_51_cast_fp16")];
+            tensor<fp16, [1024]> input_147_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_147_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(261342080)))];
+            tensor<fp16, [1024]> input_147_beta_0_to_fp16 = const()[name = tensor<string, []>("input_147_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(261344192)))];
+            tensor<fp16, []> input_147_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_147_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_147_cast_fp16 = batch_norm(beta = input_147_beta_0_to_fp16, epsilon = input_147_epsilon_0_to_fp16, gamma = input_147_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_51_cast_fp16)[name = tensor<string, []>("input_147_cast_fp16")];
+            tensor<string, []> input_149_pad_type_0 = const()[name = tensor<string, []>("input_149_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_149_strides_0 = const()[name = tensor<string, []>("input_149_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_149_pad_0 = const()[name = tensor<string, []>("input_149_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_149_dilations_0 = const()[name = tensor<string, []>("input_149_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_149_groups_0 = const()[name = tensor<string, []>("input_149_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_5_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_feed_forward1_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(261346304)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_147_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_147_dilations_0, groups = input_147_groups_0, pad = input_147_pad_0, pad_type = input_147_pad_type_0, strides = input_147_strides_0, weight = layers_5_feed_forward1_fc1_weight_to_fp16, x = input_145_cast_fp16)[name = tensor<string, []>("input_147_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_149_cast_fp16 = silu(x = input_147_cast_fp16)[name = tensor<string, []>("input_149_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_149_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = layers_5_feed_forward1_fc1_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("input_149_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_151_cast_fp16 = silu(x = input_149_cast_fp16)[name = tensor<string, []>("input_151_cast_fp16")];
             tensor<string, []> x_33_pad_type_0 = const()[name = tensor<string, []>("x_33_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_33_strides_0 = const()[name = tensor<string, []>("x_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_33_pad_0 = const()[name = tensor<string, []>("x_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_33_dilations_0 = const()[name = tensor<string, []>("x_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_33_groups_0 = const()[name = tensor<string, []>("x_33_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_1602_weight_0_to_fp16 = const()[name = tensor<string, []>("op_1602_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(269734976)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_1602_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_33_dilations_0, groups = x_33_groups_0, pad = x_33_pad_0, pad_type = x_33_pad_type_0, strides = x_33_strides_0, weight = var_1602_weight_0_to_fp16, x = input_149_cast_fp16)[name = tensor<string, []>("op_1602_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_53_cast_fp16 = add(x = inputs_51_cast_fp16, y = var_1602_cast_fp16)[name = tensor<string, []>("inputs_53_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_1600_weight_0_to_fp16 = const()[name = tensor<string, []>("op_1600_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(269734976)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_1600_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_33_dilations_0, groups = x_33_groups_0, pad = x_33_pad_0, pad_type = x_33_pad_type_0, strides = x_33_strides_0, weight = var_1600_weight_0_to_fp16, x = input_151_cast_fp16)[name = tensor<string, []>("op_1600_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_53_cast_fp16 = add(x = inputs_51_cast_fp16, y = var_1600_cast_fp16)[name = tensor<string, []>("inputs_53_cast_fp16")];
             tensor<int32, [1]> out_53_axes_0 = const()[name = tensor<string, []>("out_53_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_1612_to_fp16 = const()[name = tensor<string, []>("op_1612_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_53_cast_fp16 = layer_norm(axes = out_53_axes_0, epsilon = var_1612_to_fp16, x = inputs_53_cast_fp16)[name = tensor<string, []>("out_53_cast_fp16")];
+            tensor<fp16, []> var_1610_to_fp16 = const()[name = tensor<string, []>("op_1610_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_53_cast_fp16 = layer_norm(axes = out_53_axes_0, epsilon = var_1610_to_fp16, x = inputs_53_cast_fp16)[name = tensor<string, []>("out_53_cast_fp16")];
             tensor<fp16, [1024]> obj_23_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_23_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278123648)))];
             tensor<fp16, [1024]> obj_23_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_23_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278125760)))];
             tensor<fp16, []> obj_23_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_23_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> obj_23_cast_fp16 = batch_norm(beta = obj_23_beta_0_to_fp16, epsilon = obj_23_epsilon_0_to_fp16, gamma = obj_23_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_53_cast_fp16)[name = tensor<string, []>("obj_23_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_23_cast_fp16 = batch_norm(beta = obj_23_beta_0_to_fp16, epsilon = obj_23_epsilon_0_to_fp16, gamma = obj_23_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_53_cast_fp16)[name = tensor<string, []>("obj_23_cast_fp16")];
             tensor<string, []> query_21_pad_type_0 = const()[name = tensor<string, []>("query_21_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> query_21_strides_0 = const()[name = tensor<string, []>("query_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> query_21_pad_0 = const()[name = tensor<string, []>("query_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> query_21_dilations_0 = const()[name = tensor<string, []>("query_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> query_21_groups_0 = const()[name = tensor<string, []>("query_21_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_5_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278127872)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_21_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = query_21_dilations_0, groups = query_21_groups_0, pad = query_21_pad_0, pad_type = query_21_pad_type_0, strides = query_21_strides_0, weight = layers_5_self_attn_q_proj_weight_to_fp16, x = obj_23_cast_fp16)[name = tensor<string, []>("query_21_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> query_21_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_21_dilations_0, groups = query_21_groups_0, pad = query_21_pad_0, pad_type = query_21_pad_type_0, strides = query_21_strides_0, weight = layers_5_self_attn_q_proj_weight_to_fp16, x = obj_23_cast_fp16)[name = tensor<string, []>("query_21_cast_fp16")];
             tensor<string, []> key_11_pad_type_0 = const()[name = tensor<string, []>("key_11_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> key_11_strides_0 = const()[name = tensor<string, []>("key_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> key_11_pad_0 = const()[name = tensor<string, []>("key_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
@@ -1012,173 +1014,173 @@ program(1.0)
             tensor<int32, [2]> value_11_dilations_0 = const()[name = tensor<string, []>("value_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> value_11_groups_0 = const()[name = tensor<string, []>("value_11_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_5_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(282322304)))];
-            tensor<fp16, [1, 1024, 1, 188]> value_11_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = value_11_dilations_0, groups = value_11_groups_0, pad = value_11_pad_0, pad_type = value_11_pad_type_0, strides = value_11_strides_0, weight = layers_5_self_attn_v_proj_weight_to_fp16, x = obj_23_cast_fp16)[name = tensor<string, []>("value_11_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_1650_to_fp16 = const()[name = tensor<string, []>("op_1650_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(284419520)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_23_cast_fp16 = add(x = query_21_cast_fp16, y = var_1650_to_fp16)[name = tensor<string, []>("query_23_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_1653_to_fp16 = const()[name = tensor<string, []>("op_1653_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(284421632)))];
-            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_11_cast_fp16 = add(x = query_21_cast_fp16, y = var_1653_to_fp16)[name = tensor<string, []>("q_with_bias_v_11_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> value_11_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_11_dilations_0, groups = value_11_groups_0, pad = value_11_pad_0, pad_type = value_11_pad_type_0, strides = value_11_strides_0, weight = layers_5_self_attn_v_proj_weight_to_fp16, x = obj_23_cast_fp16)[name = tensor<string, []>("value_11_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_1648_to_fp16 = const()[name = tensor<string, []>("op_1648_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(284419520)))];
+            tensor<fp16, [1, 1024, 1, 188]> query_23_cast_fp16 = add(x = query_21_cast_fp16, y = var_1648_to_fp16)[name = tensor<string, []>("query_23_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_1651_to_fp16 = const()[name = tensor<string, []>("op_1651_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(284421632)))];
+            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_11_cast_fp16 = add(x = query_21_cast_fp16, y = var_1651_to_fp16)[name = tensor<string, []>("q_with_bias_v_11_cast_fp16")];
             tensor<string, []> p_11_pad_type_0 = const()[name = tensor<string, []>("p_11_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> p_11_strides_0 = const()[name = tensor<string, []>("p_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> p_11_pad_0 = const()[name = tensor<string, []>("p_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> p_11_dilations_0 = const()[name = tensor<string, []>("p_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> p_11_groups_0 = const()[name = tensor<string, []>("p_11_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_5_self_attn_linear_pos_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_linear_pos_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(284423744)))];
-            tensor<fp16, [1, 1024, 1, 375]> p_11_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = p_11_dilations_0, groups = p_11_groups_0, pad = p_11_pad_0, pad_type = p_11_pad_type_0, strides = p_11_strides_0, weight = layers_5_self_attn_linear_pos_weight_to_fp16, x = pos_enc_to_fp16)[name = tensor<string, []>("p_11_cast_fp16")];
-            tensor<int32, [4]> var_1664 = const()[name = tensor<string, []>("op_1664"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_1665_cast_fp16 = reshape(shape = var_1664, x = q_with_bias_v_11_cast_fp16)[name = tensor<string, []>("op_1665_cast_fp16")];
-            tensor<int32, [4]> var_1666 = const()[name = tensor<string, []>("op_1666"), val = tensor<int32, [4]>([1, 8, 128, -1])];
-            tensor<fp16, [1, 8, 128, 375]> var_1667_cast_fp16 = reshape(shape = var_1666, x = p_11_cast_fp16)[name = tensor<string, []>("op_1667_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 375]> p_11_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_11_dilations_0, groups = p_11_groups_0, pad = p_11_pad_0, pad_type = p_11_pad_type_0, strides = p_11_strides_0, weight = layers_5_self_attn_linear_pos_weight_to_fp16, x = obj_3_cast_fp16)[name = tensor<string, []>("p_11_cast_fp16")];
+            tensor<int32, [4]> var_1662 = const()[name = tensor<string, []>("op_1662"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_1663_cast_fp16 = reshape(shape = var_1662, x = q_with_bias_v_11_cast_fp16)[name = tensor<string, []>("op_1663_cast_fp16")];
+            tensor<int32, [4]> var_1664 = const()[name = tensor<string, []>("op_1664"), val = tensor<int32, [4]>([1, 8, 128, -1])];
+            tensor<fp16, [1, 8, 128, 375]> var_1665_cast_fp16 = reshape(shape = var_1664, x = p_11_cast_fp16)[name = tensor<string, []>("op_1665_cast_fp16")];
             tensor<bool, []> matrix_bd_41_transpose_x_0 = const()[name = tensor<string, []>("matrix_bd_41_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> matrix_bd_41_transpose_y_0 = const()[name = tensor<string, []>("matrix_bd_41_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_41_cast_fp16 = matmul(transpose_x = matrix_bd_41_transpose_x_0, transpose_y = matrix_bd_41_transpose_y_0, x = var_1665_cast_fp16, y = var_1667_cast_fp16)[name = tensor<string, []>("matrix_bd_41_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_41_cast_fp16 = matmul(transpose_x = matrix_bd_41_transpose_x_0, transpose_y = matrix_bd_41_transpose_y_0, x = var_1663_cast_fp16, y = var_1665_cast_fp16)[name = tensor<string, []>("matrix_bd_41_cast_fp16")];
             tensor<int32, [8]> matrix_bd_43_pad_0 = const()[name = tensor<string, []>("matrix_bd_43_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
             tensor<string, []> matrix_bd_43_mode_0 = const()[name = tensor<string, []>("matrix_bd_43_mode_0"), val = tensor<string, []>("constant")];
             tensor<fp16, []> const_65_to_fp16 = const()[name = tensor<string, []>("const_65_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
             tensor<fp16, [1, 8, 188, 376]> matrix_bd_43_cast_fp16 = pad(constant_val = const_65_to_fp16, mode = matrix_bd_43_mode_0, pad = matrix_bd_43_pad_0, x = matrix_bd_41_cast_fp16)[name = tensor<string, []>("matrix_bd_43_cast_fp16")];
-            tensor<int32, [4]> var_1676 = const()[name = tensor<string, []>("op_1676"), val = tensor<int32, [4]>([1, 8, -1, 188])];
-            tensor<fp16, [1, 8, 376, 188]> matrix_bd_45_cast_fp16 = reshape(shape = var_1676, x = matrix_bd_43_cast_fp16)[name = tensor<string, []>("matrix_bd_45_cast_fp16")];
-            tensor<int32, [4]> var_1680_begin_0 = const()[name = tensor<string, []>("op_1680_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
-            tensor<int32, [4]> var_1680_end_0 = const()[name = tensor<string, []>("op_1680_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
-            tensor<bool, [4]> var_1680_end_mask_0 = const()[name = tensor<string, []>("op_1680_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
-            tensor<fp16, [1, 8, 375, 188]> var_1680_cast_fp16 = slice_by_index(begin = var_1680_begin_0, end = var_1680_end_0, end_mask = var_1680_end_mask_0, x = matrix_bd_45_cast_fp16)[name = tensor<string, []>("op_1680_cast_fp16")];
-            tensor<int32, [4]> var_1681 = const()[name = tensor<string, []>("op_1681"), val = tensor<int32, [4]>([1, 8, 188, 375])];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_47_cast_fp16 = reshape(shape = var_1681, x = var_1680_cast_fp16)[name = tensor<string, []>("matrix_bd_47_cast_fp16")];
-            tensor<int32, [4]> var_1686_begin_0 = const()[name = tensor<string, []>("op_1686_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [4]> var_1686_end_0 = const()[name = tensor<string, []>("op_1686_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
-            tensor<bool, [4]> var_1686_end_mask_0 = const()[name = tensor<string, []>("op_1686_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
-            tensor<fp16, [1, 8, 188, 188]> var_1686_cast_fp16 = slice_by_index(begin = var_1686_begin_0, end = var_1686_end_0, end_mask = var_1686_end_mask_0, x = matrix_bd_47_cast_fp16)[name = tensor<string, []>("op_1686_cast_fp16")];
-            tensor<fp16, []> var_1687_to_fp16 = const()[name = tensor<string, []>("op_1687_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 188, 188]> qk_mask_11_cast_fp16 = mul(x = var_1686_cast_fp16, y = var_1687_to_fp16)[name = tensor<string, []>("qk_mask_11_cast_fp16")];
-            tensor<int32, [4]> var_1691 = const()[name = tensor<string, []>("op_1691"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> mh_q_11_cast_fp16 = reshape(shape = var_1691, x = query_23_cast_fp16)[name = tensor<string, []>("mh_q_11_cast_fp16")];
-            tensor<fp16, []> var_1693_to_fp16 = const()[name = tensor<string, []>("op_1693_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 128, 188]> var_1694_cast_fp16 = mul(x = mh_q_11_cast_fp16, y = var_1693_to_fp16)[name = tensor<string, []>("op_1694_cast_fp16")];
-            tensor<int32, [4]> var_1697 = const()[name = tensor<string, []>("op_1697"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_1698_cast_fp16 = reshape(shape = var_1697, x = key_11_cast_fp16)[name = tensor<string, []>("op_1698_cast_fp16")];
+            tensor<int32, [4]> var_1674 = const()[name = tensor<string, []>("op_1674"), val = tensor<int32, [4]>([1, 8, -1, 188])];
+            tensor<fp16, [1, 8, 376, 188]> matrix_bd_45_cast_fp16 = reshape(shape = var_1674, x = matrix_bd_43_cast_fp16)[name = tensor<string, []>("matrix_bd_45_cast_fp16")];
+            tensor<int32, [4]> var_1678_begin_0 = const()[name = tensor<string, []>("op_1678_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1678_end_0 = const()[name = tensor<string, []>("op_1678_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
+            tensor<bool, [4]> var_1678_end_mask_0 = const()[name = tensor<string, []>("op_1678_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 375, 188]> var_1678_cast_fp16 = slice_by_index(begin = var_1678_begin_0, end = var_1678_end_0, end_mask = var_1678_end_mask_0, x = matrix_bd_45_cast_fp16)[name = tensor<string, []>("op_1678_cast_fp16")];
+            tensor<int32, [4]> var_1679 = const()[name = tensor<string, []>("op_1679"), val = tensor<int32, [4]>([1, 8, 188, 375])];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_47_cast_fp16 = reshape(shape = var_1679, x = var_1678_cast_fp16)[name = tensor<string, []>("matrix_bd_47_cast_fp16")];
+            tensor<int32, [4]> var_1684_begin_0 = const()[name = tensor<string, []>("op_1684_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_1684_end_0 = const()[name = tensor<string, []>("op_1684_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
+            tensor<bool, [4]> var_1684_end_mask_0 = const()[name = tensor<string, []>("op_1684_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 188, 188]> var_1684_cast_fp16 = slice_by_index(begin = var_1684_begin_0, end = var_1684_end_0, end_mask = var_1684_end_mask_0, x = matrix_bd_47_cast_fp16)[name = tensor<string, []>("op_1684_cast_fp16")];
+            tensor<fp16, []> var_1685_to_fp16 = const()[name = tensor<string, []>("op_1685_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 188, 188]> qk_mask_11_cast_fp16 = mul(x = var_1684_cast_fp16, y = var_1685_to_fp16)[name = tensor<string, []>("qk_mask_11_cast_fp16")];
+            tensor<int32, [4]> var_1689 = const()[name = tensor<string, []>("op_1689"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> mh_q_11_cast_fp16 = reshape(shape = var_1689, x = query_23_cast_fp16)[name = tensor<string, []>("mh_q_11_cast_fp16")];
+            tensor<fp16, []> var_1691_to_fp16 = const()[name = tensor<string, []>("op_1691_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 128, 188]> var_1692_cast_fp16 = mul(x = mh_q_11_cast_fp16, y = var_1691_to_fp16)[name = tensor<string, []>("op_1692_cast_fp16")];
+            tensor<int32, [4]> var_1695 = const()[name = tensor<string, []>("op_1695"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_1696_cast_fp16 = reshape(shape = var_1695, x = key_11_cast_fp16)[name = tensor<string, []>("op_1696_cast_fp16")];
             tensor<bool, []> mh_w_21_transpose_x_0 = const()[name = tensor<string, []>("mh_w_21_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> mh_w_21_transpose_y_0 = const()[name = tensor<string, []>("mh_w_21_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 188]> mh_w_21_cast_fp16 = matmul(transpose_x = mh_w_21_transpose_x_0, transpose_y = mh_w_21_transpose_y_0, x = var_1694_cast_fp16, y = var_1698_cast_fp16)[name = tensor<string, []>("mh_w_21_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> mh_w_21_cast_fp16 = matmul(transpose_x = mh_w_21_transpose_x_0, transpose_y = mh_w_21_transpose_y_0, x = var_1692_cast_fp16, y = var_1696_cast_fp16)[name = tensor<string, []>("mh_w_21_cast_fp16")];
             tensor<fp16, [1, 8, 188, 188]> mh_w_23_cast_fp16 = add(x = mh_w_21_cast_fp16, y = qk_mask_11_cast_fp16)[name = tensor<string, []>("mh_w_23_cast_fp16")];
-            tensor<fp16, [1, 8, 188, 188]> var_1702_cast_fp16 = softmax(axis = var_1543, x = mh_w_23_cast_fp16)[name = tensor<string, []>("op_1702_cast_fp16")];
-            tensor<int32, [4]> var_1703 = const()[name = tensor<string, []>("op_1703"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_1704_cast_fp16 = reshape(shape = var_1703, x = value_11_cast_fp16)[name = tensor<string, []>("op_1704_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> var_1700_cast_fp16 = softmax(axis = var_1541, x = mh_w_23_cast_fp16)[name = tensor<string, []>("op_1700_cast_fp16")];
+            tensor<int32, [4]> var_1701 = const()[name = tensor<string, []>("op_1701"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_1702_cast_fp16 = reshape(shape = var_1701, x = value_11_cast_fp16)[name = tensor<string, []>("op_1702_cast_fp16")];
             tensor<bool, []> attn_11_transpose_x_0 = const()[name = tensor<string, []>("attn_11_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> attn_11_transpose_y_0 = const()[name = tensor<string, []>("attn_11_transpose_y_0"), val = tensor<bool, []>(true)];
-            tensor<fp16, [1, 8, 128, 188]> attn_11_cast_fp16 = matmul(transpose_x = attn_11_transpose_x_0, transpose_y = attn_11_transpose_y_0, x = var_1704_cast_fp16, y = var_1702_cast_fp16)[name = tensor<string, []>("attn_11_cast_fp16")];
-            tensor<int32, [4]> var_1707 = const()[name = tensor<string, []>("op_1707"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
-            tensor<fp16, [1, 1024, 1, 188]> input_151_cast_fp16 = reshape(shape = var_1707, x = attn_11_cast_fp16)[name = tensor<string, []>("input_151_cast_fp16")];
+            tensor<fp16, [1, 8, 128, 188]> attn_11_cast_fp16 = matmul(transpose_x = attn_11_transpose_x_0, transpose_y = attn_11_transpose_y_0, x = var_1702_cast_fp16, y = var_1700_cast_fp16)[name = tensor<string, []>("attn_11_cast_fp16")];
+            tensor<int32, [4]> var_1705 = const()[name = tensor<string, []>("op_1705"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
+            tensor<fp16, [1, 1024, 1, 188]> input_153_cast_fp16 = reshape(shape = var_1705, x = attn_11_cast_fp16)[name = tensor<string, []>("input_153_cast_fp16")];
             tensor<string, []> obj_25_pad_type_0 = const()[name = tensor<string, []>("obj_25_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> obj_25_strides_0 = const()[name = tensor<string, []>("obj_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> obj_25_pad_0 = const()[name = tensor<string, []>("obj_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> obj_25_dilations_0 = const()[name = tensor<string, []>("obj_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> obj_25_groups_0 = const()[name = tensor<string, []>("obj_25_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_5_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286520960)))];
-            tensor<fp16, [1, 1024, 1, 188]> obj_25_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = obj_25_dilations_0, groups = obj_25_groups_0, pad = obj_25_pad_0, pad_type = obj_25_pad_type_0, strides = obj_25_strides_0, weight = layers_5_self_attn_o_proj_weight_to_fp16, x = input_151_cast_fp16)[name = tensor<string, []>("obj_25_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_25_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_25_dilations_0, groups = obj_25_groups_0, pad = obj_25_pad_0, pad_type = obj_25_pad_type_0, strides = obj_25_strides_0, weight = layers_5_self_attn_o_proj_weight_to_fp16, x = input_153_cast_fp16)[name = tensor<string, []>("obj_25_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_55_cast_fp16 = add(x = inputs_53_cast_fp16, y = obj_25_cast_fp16)[name = tensor<string, []>("inputs_55_cast_fp16")];
             tensor<int32, [1]> out_55_axes_0 = const()[name = tensor<string, []>("out_55_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_1725_to_fp16 = const()[name = tensor<string, []>("op_1725_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_55_cast_fp16 = layer_norm(axes = out_55_axes_0, epsilon = var_1725_to_fp16, x = inputs_55_cast_fp16)[name = tensor<string, []>("out_55_cast_fp16")];
-            tensor<fp16, [1024]> input_153_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_153_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(288618176)))];
-            tensor<fp16, [1024]> input_153_beta_0_to_fp16 = const()[name = tensor<string, []>("input_153_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(288620288)))];
-            tensor<fp16, []> input_153_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_153_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_153_cast_fp16 = batch_norm(beta = input_153_beta_0_to_fp16, epsilon = input_153_epsilon_0_to_fp16, gamma = input_153_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_55_cast_fp16)[name = tensor<string, []>("input_153_cast_fp16")];
-            tensor<string, []> input_155_pad_type_0 = const()[name = tensor<string, []>("input_155_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_155_strides_0 = const()[name = tensor<string, []>("input_155_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_155_pad_0 = const()[name = tensor<string, []>("input_155_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_155_dilations_0 = const()[name = tensor<string, []>("input_155_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_155_groups_0 = const()[name = tensor<string, []>("input_155_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_1723_to_fp16 = const()[name = tensor<string, []>("op_1723_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_55_cast_fp16 = layer_norm(axes = out_55_axes_0, epsilon = var_1723_to_fp16, x = inputs_55_cast_fp16)[name = tensor<string, []>("out_55_cast_fp16")];
+            tensor<fp16, [1024]> input_155_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_155_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(288618176)))];
+            tensor<fp16, [1024]> input_155_beta_0_to_fp16 = const()[name = tensor<string, []>("input_155_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(288620288)))];
+            tensor<fp16, []> input_155_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_155_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_155_cast_fp16 = batch_norm(beta = input_155_beta_0_to_fp16, epsilon = input_155_epsilon_0_to_fp16, gamma = input_155_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_55_cast_fp16)[name = tensor<string, []>("input_155_cast_fp16")];
+            tensor<string, []> input_157_pad_type_0 = const()[name = tensor<string, []>("input_157_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_157_strides_0 = const()[name = tensor<string, []>("input_157_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_157_pad_0 = const()[name = tensor<string, []>("input_157_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_157_dilations_0 = const()[name = tensor<string, []>("input_157_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_157_groups_0 = const()[name = tensor<string, []>("input_157_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [2048, 1024, 1, 1]> layers_5_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(288622400)))];
-            tensor<fp16, [1, 2048, 1, 188]> input_155_cast_fp16 = conv(dilations = input_155_dilations_0, groups = input_155_groups_0, pad = input_155_pad_0, pad_type = input_155_pad_type_0, strides = input_155_strides_0, weight = layers_5_conv_pointwise_conv1_weight_to_fp16, x = input_153_cast_fp16)[name = tensor<string, []>("input_155_cast_fp16")];
-            tensor<int32, []> input_157_split_num_splits_0 = const()[name = tensor<string, []>("input_157_split_num_splits_0"), val = tensor<int32, []>(2)];
-            tensor<int32, []> input_157_split_axis_0 = const()[name = tensor<string, []>("input_157_split_axis_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1, 1024, 1, 188]> input_157_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_157_split_cast_fp16_1 = split(axis = input_157_split_axis_0, num_splits = input_157_split_num_splits_0, x = input_155_cast_fp16)[name = tensor<string, []>("input_157_split_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_157_split_1_sigmoid_cast_fp16 = sigmoid(x = input_157_split_cast_fp16_1)[name = tensor<string, []>("input_157_split_1_sigmoid_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_157_cast_fp16 = mul(x = input_157_split_cast_fp16_0, y = input_157_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_157_cast_fp16")];
-            tensor<string, []> input_159_pad_type_0 = const()[name = tensor<string, []>("input_159_pad_type_0"), val = tensor<string, []>("custom")];
-            tensor<int32, [4]> input_159_pad_0 = const()[name = tensor<string, []>("input_159_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
-            tensor<int32, []> input_159_groups_0 = const()[name = tensor<string, []>("input_159_groups_0"), val = tensor<int32, []>(1024)];
-            tensor<int32, [2]> input_159_strides_0 = const()[name = tensor<string, []>("input_159_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> input_159_dilations_0 = const()[name = tensor<string, []>("input_159_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [1, 2048, 1, 188]> input_157_cast_fp16 = conv(dilations = input_157_dilations_0, groups = input_157_groups_0, pad = input_157_pad_0, pad_type = input_157_pad_type_0, strides = input_157_strides_0, weight = layers_5_conv_pointwise_conv1_weight_to_fp16, x = input_155_cast_fp16)[name = tensor<string, []>("input_157_cast_fp16")];
+            tensor<int32, []> input_159_split_num_splits_0 = const()[name = tensor<string, []>("input_159_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> input_159_split_axis_0 = const()[name = tensor<string, []>("input_159_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 1, 188]> input_159_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_159_split_cast_fp16_1 = split(axis = input_159_split_axis_0, num_splits = input_159_split_num_splits_0, x = input_157_cast_fp16)[name = tensor<string, []>("input_159_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_159_split_1_sigmoid_cast_fp16 = sigmoid(x = input_159_split_cast_fp16_1)[name = tensor<string, []>("input_159_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_159_cast_fp16 = mul(x = input_159_split_cast_fp16_0, y = input_159_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_159_cast_fp16")];
+            tensor<string, []> input_161_pad_type_0 = const()[name = tensor<string, []>("input_161_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_161_pad_0 = const()[name = tensor<string, []>("input_161_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
+            tensor<int32, []> input_161_groups_0 = const()[name = tensor<string, []>("input_161_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [2]> input_161_strides_0 = const()[name = tensor<string, []>("input_161_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> input_161_dilations_0 = const()[name = tensor<string, []>("input_161_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<fp16, [1024, 1, 1, 9]> const_278_to_fp16 = const()[name = tensor<string, []>("const_278_to_fp16"), val = tensor<fp16, [1024, 1, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(292816768)))];
             tensor<fp16, [1024]> const_279_to_fp16 = const()[name = tensor<string, []>("const_279_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(292835264)))];
-            tensor<fp16, [1, 1024, 1, 188]> input_161_cast_fp16 = conv(bias = const_279_to_fp16, dilations = input_159_dilations_0, groups = input_159_groups_0, pad = input_159_pad_0, pad_type = input_159_pad_type_0, strides = input_159_strides_0, weight = const_278_to_fp16, x = input_157_cast_fp16)[name = tensor<string, []>("input_161_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_163_cast_fp16 = silu(x = input_161_cast_fp16)[name = tensor<string, []>("input_163_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_163_cast_fp16 = conv(bias = const_279_to_fp16, dilations = input_161_dilations_0, groups = input_161_groups_0, pad = input_161_pad_0, pad_type = input_161_pad_type_0, strides = input_161_strides_0, weight = const_278_to_fp16, x = input_159_cast_fp16)[name = tensor<string, []>("input_163_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_165_cast_fp16 = silu(x = input_163_cast_fp16)[name = tensor<string, []>("input_165_cast_fp16")];
             tensor<string, []> x_35_pad_type_0 = const()[name = tensor<string, []>("x_35_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_35_strides_0 = const()[name = tensor<string, []>("x_35_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_35_pad_0 = const()[name = tensor<string, []>("x_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_35_dilations_0 = const()[name = tensor<string, []>("x_35_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_35_groups_0 = const()[name = tensor<string, []>("x_35_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_5_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(292837376)))];
-            tensor<fp16, [1, 1024, 1, 188]> x_35_cast_fp16 = conv(dilations = x_35_dilations_0, groups = x_35_groups_0, pad = x_35_pad_0, pad_type = x_35_pad_type_0, strides = x_35_strides_0, weight = layers_5_conv_pointwise_conv2_weight_to_fp16, x = input_163_cast_fp16)[name = tensor<string, []>("x_35_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> x_35_cast_fp16 = conv(dilations = x_35_dilations_0, groups = x_35_groups_0, pad = x_35_pad_0, pad_type = x_35_pad_type_0, strides = x_35_strides_0, weight = layers_5_conv_pointwise_conv2_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("x_35_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_57_cast_fp16 = add(x = inputs_55_cast_fp16, y = x_35_cast_fp16)[name = tensor<string, []>("inputs_57_cast_fp16")];
             tensor<int32, [1]> out_57_axes_0 = const()[name = tensor<string, []>("out_57_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_1773_to_fp16 = const()[name = tensor<string, []>("op_1773_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_57_cast_fp16 = layer_norm(axes = out_57_axes_0, epsilon = var_1773_to_fp16, x = inputs_57_cast_fp16)[name = tensor<string, []>("out_57_cast_fp16")];
-            tensor<fp16, [1024]> input_165_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_165_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(294934592)))];
-            tensor<fp16, [1024]> input_165_beta_0_to_fp16 = const()[name = tensor<string, []>("input_165_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(294936704)))];
-            tensor<fp16, []> input_165_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_165_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_165_cast_fp16 = batch_norm(beta = input_165_beta_0_to_fp16, epsilon = input_165_epsilon_0_to_fp16, gamma = input_165_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_57_cast_fp16)[name = tensor<string, []>("input_165_cast_fp16")];
-            tensor<string, []> input_167_pad_type_0 = const()[name = tensor<string, []>("input_167_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_167_strides_0 = const()[name = tensor<string, []>("input_167_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_167_pad_0 = const()[name = tensor<string, []>("input_167_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_167_dilations_0 = const()[name = tensor<string, []>("input_167_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_167_groups_0 = const()[name = tensor<string, []>("input_167_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_1771_to_fp16 = const()[name = tensor<string, []>("op_1771_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_57_cast_fp16 = layer_norm(axes = out_57_axes_0, epsilon = var_1771_to_fp16, x = inputs_57_cast_fp16)[name = tensor<string, []>("out_57_cast_fp16")];
+            tensor<fp16, [1024]> input_167_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_167_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(294934592)))];
+            tensor<fp16, [1024]> input_167_beta_0_to_fp16 = const()[name = tensor<string, []>("input_167_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(294936704)))];
+            tensor<fp16, []> input_167_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_167_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_167_cast_fp16 = batch_norm(beta = input_167_beta_0_to_fp16, epsilon = input_167_epsilon_0_to_fp16, gamma = input_167_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_57_cast_fp16)[name = tensor<string, []>("input_167_cast_fp16")];
+            tensor<string, []> input_169_pad_type_0 = const()[name = tensor<string, []>("input_169_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_169_strides_0 = const()[name = tensor<string, []>("input_169_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_169_pad_0 = const()[name = tensor<string, []>("input_169_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_169_dilations_0 = const()[name = tensor<string, []>("input_169_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_169_groups_0 = const()[name = tensor<string, []>("input_169_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_5_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_feed_forward2_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(294938816)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_167_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_167_dilations_0, groups = input_167_groups_0, pad = input_167_pad_0, pad_type = input_167_pad_type_0, strides = input_167_strides_0, weight = layers_5_feed_forward2_fc1_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("input_167_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_169_cast_fp16 = silu(x = input_167_cast_fp16)[name = tensor<string, []>("input_169_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_169_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_169_dilations_0, groups = input_169_groups_0, pad = input_169_pad_0, pad_type = input_169_pad_type_0, strides = input_169_strides_0, weight = layers_5_feed_forward2_fc1_weight_to_fp16, x = input_167_cast_fp16)[name = tensor<string, []>("input_169_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_171_cast_fp16 = silu(x = input_169_cast_fp16)[name = tensor<string, []>("input_171_cast_fp16")];
             tensor<string, []> x_37_pad_type_0 = const()[name = tensor<string, []>("x_37_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_37_strides_0 = const()[name = tensor<string, []>("x_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_37_pad_0 = const()[name = tensor<string, []>("x_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_37_dilations_0 = const()[name = tensor<string, []>("x_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_37_groups_0 = const()[name = tensor<string, []>("x_37_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_1801_weight_0_to_fp16 = const()[name = tensor<string, []>("op_1801_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(303327488)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_1801_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_37_dilations_0, groups = x_37_groups_0, pad = x_37_pad_0, pad_type = x_37_pad_type_0, strides = x_37_strides_0, weight = var_1801_weight_0_to_fp16, x = input_169_cast_fp16)[name = tensor<string, []>("op_1801_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_59_cast_fp16 = add(x = inputs_57_cast_fp16, y = var_1801_cast_fp16)[name = tensor<string, []>("inputs_59_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_1799_weight_0_to_fp16 = const()[name = tensor<string, []>("op_1799_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(303327488)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_1799_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_37_dilations_0, groups = x_37_groups_0, pad = x_37_pad_0, pad_type = x_37_pad_type_0, strides = x_37_strides_0, weight = var_1799_weight_0_to_fp16, x = input_171_cast_fp16)[name = tensor<string, []>("op_1799_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_59_cast_fp16 = add(x = inputs_57_cast_fp16, y = var_1799_cast_fp16)[name = tensor<string, []>("inputs_59_cast_fp16")];
             tensor<int32, [1]> out_59_axes_0 = const()[name = tensor<string, []>("out_59_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_1811_to_fp16 = const()[name = tensor<string, []>("op_1811_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_59_cast_fp16 = layer_norm(axes = out_59_axes_0, epsilon = var_1811_to_fp16, x = inputs_59_cast_fp16)[name = tensor<string, []>("out_59_cast_fp16")];
+            tensor<fp16, []> var_1809_to_fp16 = const()[name = tensor<string, []>("op_1809_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_59_cast_fp16 = layer_norm(axes = out_59_axes_0, epsilon = var_1809_to_fp16, x = inputs_59_cast_fp16)[name = tensor<string, []>("out_59_cast_fp16")];
             tensor<fp16, [1024]> inputs_61_gamma_0_to_fp16 = const()[name = tensor<string, []>("inputs_61_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(311716160)))];
             tensor<fp16, [1024]> inputs_61_beta_0_to_fp16 = const()[name = tensor<string, []>("inputs_61_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(311718272)))];
             tensor<fp16, []> inputs_61_epsilon_0_to_fp16 = const()[name = tensor<string, []>("inputs_61_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_61_cast_fp16 = batch_norm(beta = inputs_61_beta_0_to_fp16, epsilon = inputs_61_epsilon_0_to_fp16, gamma = inputs_61_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_59_cast_fp16)[name = tensor<string, []>("inputs_61_cast_fp16")];
-            tensor<int32, []> var_1825 = const()[name = tensor<string, []>("op_1825"), val = tensor<int32, []>(3)];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_61_cast_fp16 = batch_norm(beta = inputs_61_beta_0_to_fp16, epsilon = inputs_61_epsilon_0_to_fp16, gamma = inputs_61_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_59_cast_fp16)[name = tensor<string, []>("inputs_61_cast_fp16")];
+            tensor<int32, []> var_1823 = const()[name = tensor<string, []>("op_1823"), val = tensor<int32, []>(3)];
             tensor<int32, [1]> out_61_axes_0 = const()[name = tensor<string, []>("out_61_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_1856_to_fp16 = const()[name = tensor<string, []>("op_1856_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_61_cast_fp16 = layer_norm(axes = out_61_axes_0, epsilon = var_1856_to_fp16, x = inputs_61_cast_fp16)[name = tensor<string, []>("out_61_cast_fp16")];
-            tensor<fp16, [1024]> input_171_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_171_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(311720384)))];
-            tensor<fp16, [1024]> input_171_beta_0_to_fp16 = const()[name = tensor<string, []>("input_171_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(311722496)))];
-            tensor<fp16, []> input_171_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_171_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_171_cast_fp16 = batch_norm(beta = input_171_beta_0_to_fp16, epsilon = input_171_epsilon_0_to_fp16, gamma = input_171_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_61_cast_fp16)[name = tensor<string, []>("input_171_cast_fp16")];
-            tensor<string, []> input_173_pad_type_0 = const()[name = tensor<string, []>("input_173_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_173_strides_0 = const()[name = tensor<string, []>("input_173_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_173_pad_0 = const()[name = tensor<string, []>("input_173_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_173_dilations_0 = const()[name = tensor<string, []>("input_173_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_173_groups_0 = const()[name = tensor<string, []>("input_173_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_1854_to_fp16 = const()[name = tensor<string, []>("op_1854_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_61_cast_fp16 = layer_norm(axes = out_61_axes_0, epsilon = var_1854_to_fp16, x = inputs_61_cast_fp16)[name = tensor<string, []>("out_61_cast_fp16")];
+            tensor<fp16, [1024]> input_173_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_173_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(311720384)))];
+            tensor<fp16, [1024]> input_173_beta_0_to_fp16 = const()[name = tensor<string, []>("input_173_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(311722496)))];
+            tensor<fp16, []> input_173_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_173_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_173_cast_fp16 = batch_norm(beta = input_173_beta_0_to_fp16, epsilon = input_173_epsilon_0_to_fp16, gamma = input_173_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_61_cast_fp16)[name = tensor<string, []>("input_173_cast_fp16")];
+            tensor<string, []> input_175_pad_type_0 = const()[name = tensor<string, []>("input_175_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_175_strides_0 = const()[name = tensor<string, []>("input_175_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_175_pad_0 = const()[name = tensor<string, []>("input_175_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_175_dilations_0 = const()[name = tensor<string, []>("input_175_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_175_groups_0 = const()[name = tensor<string, []>("input_175_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_6_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_feed_forward1_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(311724608)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_173_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_173_dilations_0, groups = input_173_groups_0, pad = input_173_pad_0, pad_type = input_173_pad_type_0, strides = input_173_strides_0, weight = layers_6_feed_forward1_fc1_weight_to_fp16, x = input_171_cast_fp16)[name = tensor<string, []>("input_173_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_175_cast_fp16 = silu(x = input_173_cast_fp16)[name = tensor<string, []>("input_175_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_175_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_175_dilations_0, groups = input_175_groups_0, pad = input_175_pad_0, pad_type = input_175_pad_type_0, strides = input_175_strides_0, weight = layers_6_feed_forward1_fc1_weight_to_fp16, x = input_173_cast_fp16)[name = tensor<string, []>("input_175_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_177_cast_fp16 = silu(x = input_175_cast_fp16)[name = tensor<string, []>("input_177_cast_fp16")];
             tensor<string, []> x_39_pad_type_0 = const()[name = tensor<string, []>("x_39_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_39_strides_0 = const()[name = tensor<string, []>("x_39_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_39_pad_0 = const()[name = tensor<string, []>("x_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_39_dilations_0 = const()[name = tensor<string, []>("x_39_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_39_groups_0 = const()[name = tensor<string, []>("x_39_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_1884_weight_0_to_fp16 = const()[name = tensor<string, []>("op_1884_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(320113280)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_1884_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_39_dilations_0, groups = x_39_groups_0, pad = x_39_pad_0, pad_type = x_39_pad_type_0, strides = x_39_strides_0, weight = var_1884_weight_0_to_fp16, x = input_175_cast_fp16)[name = tensor<string, []>("op_1884_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_63_cast_fp16 = add(x = inputs_61_cast_fp16, y = var_1884_cast_fp16)[name = tensor<string, []>("inputs_63_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_1882_weight_0_to_fp16 = const()[name = tensor<string, []>("op_1882_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(320113280)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_1882_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_39_dilations_0, groups = x_39_groups_0, pad = x_39_pad_0, pad_type = x_39_pad_type_0, strides = x_39_strides_0, weight = var_1882_weight_0_to_fp16, x = input_177_cast_fp16)[name = tensor<string, []>("op_1882_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_63_cast_fp16 = add(x = inputs_61_cast_fp16, y = var_1882_cast_fp16)[name = tensor<string, []>("inputs_63_cast_fp16")];
             tensor<int32, [1]> out_63_axes_0 = const()[name = tensor<string, []>("out_63_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_1894_to_fp16 = const()[name = tensor<string, []>("op_1894_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_63_cast_fp16 = layer_norm(axes = out_63_axes_0, epsilon = var_1894_to_fp16, x = inputs_63_cast_fp16)[name = tensor<string, []>("out_63_cast_fp16")];
+            tensor<fp16, []> var_1892_to_fp16 = const()[name = tensor<string, []>("op_1892_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_63_cast_fp16 = layer_norm(axes = out_63_axes_0, epsilon = var_1892_to_fp16, x = inputs_63_cast_fp16)[name = tensor<string, []>("out_63_cast_fp16")];
             tensor<fp16, [1024]> obj_27_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_27_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(328501952)))];
             tensor<fp16, [1024]> obj_27_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_27_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(328504064)))];
             tensor<fp16, []> obj_27_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_27_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> obj_27_cast_fp16 = batch_norm(beta = obj_27_beta_0_to_fp16, epsilon = obj_27_epsilon_0_to_fp16, gamma = obj_27_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_63_cast_fp16)[name = tensor<string, []>("obj_27_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_27_cast_fp16 = batch_norm(beta = obj_27_beta_0_to_fp16, epsilon = obj_27_epsilon_0_to_fp16, gamma = obj_27_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_63_cast_fp16)[name = tensor<string, []>("obj_27_cast_fp16")];
             tensor<string, []> query_25_pad_type_0 = const()[name = tensor<string, []>("query_25_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> query_25_strides_0 = const()[name = tensor<string, []>("query_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> query_25_pad_0 = const()[name = tensor<string, []>("query_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> query_25_dilations_0 = const()[name = tensor<string, []>("query_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> query_25_groups_0 = const()[name = tensor<string, []>("query_25_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_6_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(328506176)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_25_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = query_25_dilations_0, groups = query_25_groups_0, pad = query_25_pad_0, pad_type = query_25_pad_type_0, strides = query_25_strides_0, weight = layers_6_self_attn_q_proj_weight_to_fp16, x = obj_27_cast_fp16)[name = tensor<string, []>("query_25_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> query_25_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_25_dilations_0, groups = query_25_groups_0, pad = query_25_pad_0, pad_type = query_25_pad_type_0, strides = query_25_strides_0, weight = layers_6_self_attn_q_proj_weight_to_fp16, x = obj_27_cast_fp16)[name = tensor<string, []>("query_25_cast_fp16")];
             tensor<string, []> key_13_pad_type_0 = const()[name = tensor<string, []>("key_13_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> key_13_strides_0 = const()[name = tensor<string, []>("key_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> key_13_pad_0 = const()[name = tensor<string, []>("key_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
@@ -1192,173 +1194,173 @@ program(1.0)
             tensor<int32, [2]> value_13_dilations_0 = const()[name = tensor<string, []>("value_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> value_13_groups_0 = const()[name = tensor<string, []>("value_13_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_6_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(332700608)))];
-            tensor<fp16, [1, 1024, 1, 188]> value_13_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = value_13_dilations_0, groups = value_13_groups_0, pad = value_13_pad_0, pad_type = value_13_pad_type_0, strides = value_13_strides_0, weight = layers_6_self_attn_v_proj_weight_to_fp16, x = obj_27_cast_fp16)[name = tensor<string, []>("value_13_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_1932_to_fp16 = const()[name = tensor<string, []>("op_1932_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(334797824)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_27_cast_fp16 = add(x = query_25_cast_fp16, y = var_1932_to_fp16)[name = tensor<string, []>("query_27_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_1935_to_fp16 = const()[name = tensor<string, []>("op_1935_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(334799936)))];
-            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_13_cast_fp16 = add(x = query_25_cast_fp16, y = var_1935_to_fp16)[name = tensor<string, []>("q_with_bias_v_13_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> value_13_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_13_dilations_0, groups = value_13_groups_0, pad = value_13_pad_0, pad_type = value_13_pad_type_0, strides = value_13_strides_0, weight = layers_6_self_attn_v_proj_weight_to_fp16, x = obj_27_cast_fp16)[name = tensor<string, []>("value_13_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_1930_to_fp16 = const()[name = tensor<string, []>("op_1930_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(334797824)))];
+            tensor<fp16, [1, 1024, 1, 188]> query_27_cast_fp16 = add(x = query_25_cast_fp16, y = var_1930_to_fp16)[name = tensor<string, []>("query_27_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_1933_to_fp16 = const()[name = tensor<string, []>("op_1933_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(334799936)))];
+            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_13_cast_fp16 = add(x = query_25_cast_fp16, y = var_1933_to_fp16)[name = tensor<string, []>("q_with_bias_v_13_cast_fp16")];
             tensor<string, []> p_13_pad_type_0 = const()[name = tensor<string, []>("p_13_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> p_13_strides_0 = const()[name = tensor<string, []>("p_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> p_13_pad_0 = const()[name = tensor<string, []>("p_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> p_13_dilations_0 = const()[name = tensor<string, []>("p_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> p_13_groups_0 = const()[name = tensor<string, []>("p_13_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_6_self_attn_linear_pos_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_linear_pos_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(334802048)))];
-            tensor<fp16, [1, 1024, 1, 375]> p_13_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = p_13_dilations_0, groups = p_13_groups_0, pad = p_13_pad_0, pad_type = p_13_pad_type_0, strides = p_13_strides_0, weight = layers_6_self_attn_linear_pos_weight_to_fp16, x = pos_enc_to_fp16)[name = tensor<string, []>("p_13_cast_fp16")];
-            tensor<int32, [4]> var_1946 = const()[name = tensor<string, []>("op_1946"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_1947_cast_fp16 = reshape(shape = var_1946, x = q_with_bias_v_13_cast_fp16)[name = tensor<string, []>("op_1947_cast_fp16")];
-            tensor<int32, [4]> var_1948 = const()[name = tensor<string, []>("op_1948"), val = tensor<int32, [4]>([1, 8, 128, -1])];
-            tensor<fp16, [1, 8, 128, 375]> var_1949_cast_fp16 = reshape(shape = var_1948, x = p_13_cast_fp16)[name = tensor<string, []>("op_1949_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 375]> p_13_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_13_dilations_0, groups = p_13_groups_0, pad = p_13_pad_0, pad_type = p_13_pad_type_0, strides = p_13_strides_0, weight = layers_6_self_attn_linear_pos_weight_to_fp16, x = obj_3_cast_fp16)[name = tensor<string, []>("p_13_cast_fp16")];
+            tensor<int32, [4]> var_1944 = const()[name = tensor<string, []>("op_1944"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_1945_cast_fp16 = reshape(shape = var_1944, x = q_with_bias_v_13_cast_fp16)[name = tensor<string, []>("op_1945_cast_fp16")];
+            tensor<int32, [4]> var_1946 = const()[name = tensor<string, []>("op_1946"), val = tensor<int32, [4]>([1, 8, 128, -1])];
+            tensor<fp16, [1, 8, 128, 375]> var_1947_cast_fp16 = reshape(shape = var_1946, x = p_13_cast_fp16)[name = tensor<string, []>("op_1947_cast_fp16")];
             tensor<bool, []> matrix_bd_49_transpose_x_0 = const()[name = tensor<string, []>("matrix_bd_49_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> matrix_bd_49_transpose_y_0 = const()[name = tensor<string, []>("matrix_bd_49_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_49_cast_fp16 = matmul(transpose_x = matrix_bd_49_transpose_x_0, transpose_y = matrix_bd_49_transpose_y_0, x = var_1947_cast_fp16, y = var_1949_cast_fp16)[name = tensor<string, []>("matrix_bd_49_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_49_cast_fp16 = matmul(transpose_x = matrix_bd_49_transpose_x_0, transpose_y = matrix_bd_49_transpose_y_0, x = var_1945_cast_fp16, y = var_1947_cast_fp16)[name = tensor<string, []>("matrix_bd_49_cast_fp16")];
             tensor<int32, [8]> matrix_bd_51_pad_0 = const()[name = tensor<string, []>("matrix_bd_51_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
             tensor<string, []> matrix_bd_51_mode_0 = const()[name = tensor<string, []>("matrix_bd_51_mode_0"), val = tensor<string, []>("constant")];
             tensor<fp16, []> const_76_to_fp16 = const()[name = tensor<string, []>("const_76_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
             tensor<fp16, [1, 8, 188, 376]> matrix_bd_51_cast_fp16 = pad(constant_val = const_76_to_fp16, mode = matrix_bd_51_mode_0, pad = matrix_bd_51_pad_0, x = matrix_bd_49_cast_fp16)[name = tensor<string, []>("matrix_bd_51_cast_fp16")];
-            tensor<int32, [4]> var_1958 = const()[name = tensor<string, []>("op_1958"), val = tensor<int32, [4]>([1, 8, -1, 188])];
-            tensor<fp16, [1, 8, 376, 188]> matrix_bd_53_cast_fp16 = reshape(shape = var_1958, x = matrix_bd_51_cast_fp16)[name = tensor<string, []>("matrix_bd_53_cast_fp16")];
-            tensor<int32, [4]> var_1962_begin_0 = const()[name = tensor<string, []>("op_1962_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
-            tensor<int32, [4]> var_1962_end_0 = const()[name = tensor<string, []>("op_1962_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
-            tensor<bool, [4]> var_1962_end_mask_0 = const()[name = tensor<string, []>("op_1962_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
-            tensor<fp16, [1, 8, 375, 188]> var_1962_cast_fp16 = slice_by_index(begin = var_1962_begin_0, end = var_1962_end_0, end_mask = var_1962_end_mask_0, x = matrix_bd_53_cast_fp16)[name = tensor<string, []>("op_1962_cast_fp16")];
-            tensor<int32, [4]> var_1963 = const()[name = tensor<string, []>("op_1963"), val = tensor<int32, [4]>([1, 8, 188, 375])];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_55_cast_fp16 = reshape(shape = var_1963, x = var_1962_cast_fp16)[name = tensor<string, []>("matrix_bd_55_cast_fp16")];
-            tensor<int32, [4]> var_1968_begin_0 = const()[name = tensor<string, []>("op_1968_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [4]> var_1968_end_0 = const()[name = tensor<string, []>("op_1968_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
-            tensor<bool, [4]> var_1968_end_mask_0 = const()[name = tensor<string, []>("op_1968_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
-            tensor<fp16, [1, 8, 188, 188]> var_1968_cast_fp16 = slice_by_index(begin = var_1968_begin_0, end = var_1968_end_0, end_mask = var_1968_end_mask_0, x = matrix_bd_55_cast_fp16)[name = tensor<string, []>("op_1968_cast_fp16")];
-            tensor<fp16, []> var_1969_to_fp16 = const()[name = tensor<string, []>("op_1969_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 188, 188]> qk_mask_13_cast_fp16 = mul(x = var_1968_cast_fp16, y = var_1969_to_fp16)[name = tensor<string, []>("qk_mask_13_cast_fp16")];
-            tensor<int32, [4]> var_1973 = const()[name = tensor<string, []>("op_1973"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> mh_q_13_cast_fp16 = reshape(shape = var_1973, x = query_27_cast_fp16)[name = tensor<string, []>("mh_q_13_cast_fp16")];
-            tensor<fp16, []> var_1975_to_fp16 = const()[name = tensor<string, []>("op_1975_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 128, 188]> var_1976_cast_fp16 = mul(x = mh_q_13_cast_fp16, y = var_1975_to_fp16)[name = tensor<string, []>("op_1976_cast_fp16")];
-            tensor<int32, [4]> var_1979 = const()[name = tensor<string, []>("op_1979"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_1980_cast_fp16 = reshape(shape = var_1979, x = key_13_cast_fp16)[name = tensor<string, []>("op_1980_cast_fp16")];
+            tensor<int32, [4]> var_1956 = const()[name = tensor<string, []>("op_1956"), val = tensor<int32, [4]>([1, 8, -1, 188])];
+            tensor<fp16, [1, 8, 376, 188]> matrix_bd_53_cast_fp16 = reshape(shape = var_1956, x = matrix_bd_51_cast_fp16)[name = tensor<string, []>("matrix_bd_53_cast_fp16")];
+            tensor<int32, [4]> var_1960_begin_0 = const()[name = tensor<string, []>("op_1960_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1960_end_0 = const()[name = tensor<string, []>("op_1960_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
+            tensor<bool, [4]> var_1960_end_mask_0 = const()[name = tensor<string, []>("op_1960_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 375, 188]> var_1960_cast_fp16 = slice_by_index(begin = var_1960_begin_0, end = var_1960_end_0, end_mask = var_1960_end_mask_0, x = matrix_bd_53_cast_fp16)[name = tensor<string, []>("op_1960_cast_fp16")];
+            tensor<int32, [4]> var_1961 = const()[name = tensor<string, []>("op_1961"), val = tensor<int32, [4]>([1, 8, 188, 375])];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_55_cast_fp16 = reshape(shape = var_1961, x = var_1960_cast_fp16)[name = tensor<string, []>("matrix_bd_55_cast_fp16")];
+            tensor<int32, [4]> var_1966_begin_0 = const()[name = tensor<string, []>("op_1966_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_1966_end_0 = const()[name = tensor<string, []>("op_1966_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
+            tensor<bool, [4]> var_1966_end_mask_0 = const()[name = tensor<string, []>("op_1966_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 188, 188]> var_1966_cast_fp16 = slice_by_index(begin = var_1966_begin_0, end = var_1966_end_0, end_mask = var_1966_end_mask_0, x = matrix_bd_55_cast_fp16)[name = tensor<string, []>("op_1966_cast_fp16")];
+            tensor<fp16, []> var_1967_to_fp16 = const()[name = tensor<string, []>("op_1967_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 188, 188]> qk_mask_13_cast_fp16 = mul(x = var_1966_cast_fp16, y = var_1967_to_fp16)[name = tensor<string, []>("qk_mask_13_cast_fp16")];
+            tensor<int32, [4]> var_1971 = const()[name = tensor<string, []>("op_1971"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> mh_q_13_cast_fp16 = reshape(shape = var_1971, x = query_27_cast_fp16)[name = tensor<string, []>("mh_q_13_cast_fp16")];
+            tensor<fp16, []> var_1973_to_fp16 = const()[name = tensor<string, []>("op_1973_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 128, 188]> var_1974_cast_fp16 = mul(x = mh_q_13_cast_fp16, y = var_1973_to_fp16)[name = tensor<string, []>("op_1974_cast_fp16")];
+            tensor<int32, [4]> var_1977 = const()[name = tensor<string, []>("op_1977"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_1978_cast_fp16 = reshape(shape = var_1977, x = key_13_cast_fp16)[name = tensor<string, []>("op_1978_cast_fp16")];
             tensor<bool, []> mh_w_25_transpose_x_0 = const()[name = tensor<string, []>("mh_w_25_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> mh_w_25_transpose_y_0 = const()[name = tensor<string, []>("mh_w_25_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 188]> mh_w_25_cast_fp16 = matmul(transpose_x = mh_w_25_transpose_x_0, transpose_y = mh_w_25_transpose_y_0, x = var_1976_cast_fp16, y = var_1980_cast_fp16)[name = tensor<string, []>("mh_w_25_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> mh_w_25_cast_fp16 = matmul(transpose_x = mh_w_25_transpose_x_0, transpose_y = mh_w_25_transpose_y_0, x = var_1974_cast_fp16, y = var_1978_cast_fp16)[name = tensor<string, []>("mh_w_25_cast_fp16")];
             tensor<fp16, [1, 8, 188, 188]> mh_w_27_cast_fp16 = add(x = mh_w_25_cast_fp16, y = qk_mask_13_cast_fp16)[name = tensor<string, []>("mh_w_27_cast_fp16")];
-            tensor<fp16, [1, 8, 188, 188]> var_1984_cast_fp16 = softmax(axis = var_1825, x = mh_w_27_cast_fp16)[name = tensor<string, []>("op_1984_cast_fp16")];
-            tensor<int32, [4]> var_1985 = const()[name = tensor<string, []>("op_1985"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_1986_cast_fp16 = reshape(shape = var_1985, x = value_13_cast_fp16)[name = tensor<string, []>("op_1986_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> var_1982_cast_fp16 = softmax(axis = var_1823, x = mh_w_27_cast_fp16)[name = tensor<string, []>("op_1982_cast_fp16")];
+            tensor<int32, [4]> var_1983 = const()[name = tensor<string, []>("op_1983"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_1984_cast_fp16 = reshape(shape = var_1983, x = value_13_cast_fp16)[name = tensor<string, []>("op_1984_cast_fp16")];
             tensor<bool, []> attn_13_transpose_x_0 = const()[name = tensor<string, []>("attn_13_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> attn_13_transpose_y_0 = const()[name = tensor<string, []>("attn_13_transpose_y_0"), val = tensor<bool, []>(true)];
-            tensor<fp16, [1, 8, 128, 188]> attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = var_1986_cast_fp16, y = var_1984_cast_fp16)[name = tensor<string, []>("attn_13_cast_fp16")];
-            tensor<int32, [4]> var_1989 = const()[name = tensor<string, []>("op_1989"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
-            tensor<fp16, [1, 1024, 1, 188]> input_177_cast_fp16 = reshape(shape = var_1989, x = attn_13_cast_fp16)[name = tensor<string, []>("input_177_cast_fp16")];
+            tensor<fp16, [1, 8, 128, 188]> attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = var_1984_cast_fp16, y = var_1982_cast_fp16)[name = tensor<string, []>("attn_13_cast_fp16")];
+            tensor<int32, [4]> var_1987 = const()[name = tensor<string, []>("op_1987"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
+            tensor<fp16, [1, 1024, 1, 188]> input_179_cast_fp16 = reshape(shape = var_1987, x = attn_13_cast_fp16)[name = tensor<string, []>("input_179_cast_fp16")];
             tensor<string, []> obj_29_pad_type_0 = const()[name = tensor<string, []>("obj_29_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> obj_29_strides_0 = const()[name = tensor<string, []>("obj_29_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> obj_29_pad_0 = const()[name = tensor<string, []>("obj_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> obj_29_dilations_0 = const()[name = tensor<string, []>("obj_29_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> obj_29_groups_0 = const()[name = tensor<string, []>("obj_29_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_6_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(336899264)))];
-            tensor<fp16, [1, 1024, 1, 188]> obj_29_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = obj_29_dilations_0, groups = obj_29_groups_0, pad = obj_29_pad_0, pad_type = obj_29_pad_type_0, strides = obj_29_strides_0, weight = layers_6_self_attn_o_proj_weight_to_fp16, x = input_177_cast_fp16)[name = tensor<string, []>("obj_29_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_29_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_29_dilations_0, groups = obj_29_groups_0, pad = obj_29_pad_0, pad_type = obj_29_pad_type_0, strides = obj_29_strides_0, weight = layers_6_self_attn_o_proj_weight_to_fp16, x = input_179_cast_fp16)[name = tensor<string, []>("obj_29_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_65_cast_fp16 = add(x = inputs_63_cast_fp16, y = obj_29_cast_fp16)[name = tensor<string, []>("inputs_65_cast_fp16")];
             tensor<int32, [1]> out_65_axes_0 = const()[name = tensor<string, []>("out_65_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_2007_to_fp16 = const()[name = tensor<string, []>("op_2007_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_65_cast_fp16 = layer_norm(axes = out_65_axes_0, epsilon = var_2007_to_fp16, x = inputs_65_cast_fp16)[name = tensor<string, []>("out_65_cast_fp16")];
-            tensor<fp16, [1024]> input_179_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_179_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(338996480)))];
-            tensor<fp16, [1024]> input_179_beta_0_to_fp16 = const()[name = tensor<string, []>("input_179_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(338998592)))];
-            tensor<fp16, []> input_179_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_179_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_179_cast_fp16 = batch_norm(beta = input_179_beta_0_to_fp16, epsilon = input_179_epsilon_0_to_fp16, gamma = input_179_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_65_cast_fp16)[name = tensor<string, []>("input_179_cast_fp16")];
-            tensor<string, []> input_181_pad_type_0 = const()[name = tensor<string, []>("input_181_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_181_strides_0 = const()[name = tensor<string, []>("input_181_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_181_pad_0 = const()[name = tensor<string, []>("input_181_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_181_dilations_0 = const()[name = tensor<string, []>("input_181_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_181_groups_0 = const()[name = tensor<string, []>("input_181_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_2005_to_fp16 = const()[name = tensor<string, []>("op_2005_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_65_cast_fp16 = layer_norm(axes = out_65_axes_0, epsilon = var_2005_to_fp16, x = inputs_65_cast_fp16)[name = tensor<string, []>("out_65_cast_fp16")];
+            tensor<fp16, [1024]> input_181_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_181_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(338996480)))];
+            tensor<fp16, [1024]> input_181_beta_0_to_fp16 = const()[name = tensor<string, []>("input_181_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(338998592)))];
+            tensor<fp16, []> input_181_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_181_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_181_cast_fp16 = batch_norm(beta = input_181_beta_0_to_fp16, epsilon = input_181_epsilon_0_to_fp16, gamma = input_181_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_65_cast_fp16)[name = tensor<string, []>("input_181_cast_fp16")];
+            tensor<string, []> input_183_pad_type_0 = const()[name = tensor<string, []>("input_183_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_183_strides_0 = const()[name = tensor<string, []>("input_183_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_183_pad_0 = const()[name = tensor<string, []>("input_183_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_183_dilations_0 = const()[name = tensor<string, []>("input_183_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_183_groups_0 = const()[name = tensor<string, []>("input_183_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [2048, 1024, 1, 1]> layers_6_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(339000704)))];
-            tensor<fp16, [1, 2048, 1, 188]> input_181_cast_fp16 = conv(dilations = input_181_dilations_0, groups = input_181_groups_0, pad = input_181_pad_0, pad_type = input_181_pad_type_0, strides = input_181_strides_0, weight = layers_6_conv_pointwise_conv1_weight_to_fp16, x = input_179_cast_fp16)[name = tensor<string, []>("input_181_cast_fp16")];
-            tensor<int32, []> input_183_split_num_splits_0 = const()[name = tensor<string, []>("input_183_split_num_splits_0"), val = tensor<int32, []>(2)];
-            tensor<int32, []> input_183_split_axis_0 = const()[name = tensor<string, []>("input_183_split_axis_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1, 1024, 1, 188]> input_183_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_183_split_cast_fp16_1 = split(axis = input_183_split_axis_0, num_splits = input_183_split_num_splits_0, x = input_181_cast_fp16)[name = tensor<string, []>("input_183_split_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_183_split_1_sigmoid_cast_fp16 = sigmoid(x = input_183_split_cast_fp16_1)[name = tensor<string, []>("input_183_split_1_sigmoid_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_183_cast_fp16 = mul(x = input_183_split_cast_fp16_0, y = input_183_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_183_cast_fp16")];
-            tensor<string, []> input_185_pad_type_0 = const()[name = tensor<string, []>("input_185_pad_type_0"), val = tensor<string, []>("custom")];
-            tensor<int32, [4]> input_185_pad_0 = const()[name = tensor<string, []>("input_185_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
-            tensor<int32, []> input_185_groups_0 = const()[name = tensor<string, []>("input_185_groups_0"), val = tensor<int32, []>(1024)];
-            tensor<int32, [2]> input_185_strides_0 = const()[name = tensor<string, []>("input_185_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> input_185_dilations_0 = const()[name = tensor<string, []>("input_185_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [1, 2048, 1, 188]> input_183_cast_fp16 = conv(dilations = input_183_dilations_0, groups = input_183_groups_0, pad = input_183_pad_0, pad_type = input_183_pad_type_0, strides = input_183_strides_0, weight = layers_6_conv_pointwise_conv1_weight_to_fp16, x = input_181_cast_fp16)[name = tensor<string, []>("input_183_cast_fp16")];
+            tensor<int32, []> input_185_split_num_splits_0 = const()[name = tensor<string, []>("input_185_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> input_185_split_axis_0 = const()[name = tensor<string, []>("input_185_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 1, 188]> input_185_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_185_split_cast_fp16_1 = split(axis = input_185_split_axis_0, num_splits = input_185_split_num_splits_0, x = input_183_cast_fp16)[name = tensor<string, []>("input_185_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_185_split_1_sigmoid_cast_fp16 = sigmoid(x = input_185_split_cast_fp16_1)[name = tensor<string, []>("input_185_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_185_cast_fp16 = mul(x = input_185_split_cast_fp16_0, y = input_185_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_185_cast_fp16")];
+            tensor<string, []> input_187_pad_type_0 = const()[name = tensor<string, []>("input_187_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_187_pad_0 = const()[name = tensor<string, []>("input_187_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
+            tensor<int32, []> input_187_groups_0 = const()[name = tensor<string, []>("input_187_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [2]> input_187_strides_0 = const()[name = tensor<string, []>("input_187_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> input_187_dilations_0 = const()[name = tensor<string, []>("input_187_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<fp16, [1024, 1, 1, 9]> const_280_to_fp16 = const()[name = tensor<string, []>("const_280_to_fp16"), val = tensor<fp16, [1024, 1, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(343195072)))];
             tensor<fp16, [1024]> const_281_to_fp16 = const()[name = tensor<string, []>("const_281_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(343213568)))];
-            tensor<fp16, [1, 1024, 1, 188]> input_187_cast_fp16 = conv(bias = const_281_to_fp16, dilations = input_185_dilations_0, groups = input_185_groups_0, pad = input_185_pad_0, pad_type = input_185_pad_type_0, strides = input_185_strides_0, weight = const_280_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("input_187_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_189_cast_fp16 = silu(x = input_187_cast_fp16)[name = tensor<string, []>("input_189_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_189_cast_fp16 = conv(bias = const_281_to_fp16, dilations = input_187_dilations_0, groups = input_187_groups_0, pad = input_187_pad_0, pad_type = input_187_pad_type_0, strides = input_187_strides_0, weight = const_280_to_fp16, x = input_185_cast_fp16)[name = tensor<string, []>("input_189_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_191_cast_fp16 = silu(x = input_189_cast_fp16)[name = tensor<string, []>("input_191_cast_fp16")];
             tensor<string, []> x_41_pad_type_0 = const()[name = tensor<string, []>("x_41_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_41_strides_0 = const()[name = tensor<string, []>("x_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_41_pad_0 = const()[name = tensor<string, []>("x_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_41_dilations_0 = const()[name = tensor<string, []>("x_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_41_groups_0 = const()[name = tensor<string, []>("x_41_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_6_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(343215680)))];
-            tensor<fp16, [1, 1024, 1, 188]> x_41_cast_fp16 = conv(dilations = x_41_dilations_0, groups = x_41_groups_0, pad = x_41_pad_0, pad_type = x_41_pad_type_0, strides = x_41_strides_0, weight = layers_6_conv_pointwise_conv2_weight_to_fp16, x = input_189_cast_fp16)[name = tensor<string, []>("x_41_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> x_41_cast_fp16 = conv(dilations = x_41_dilations_0, groups = x_41_groups_0, pad = x_41_pad_0, pad_type = x_41_pad_type_0, strides = x_41_strides_0, weight = layers_6_conv_pointwise_conv2_weight_to_fp16, x = input_191_cast_fp16)[name = tensor<string, []>("x_41_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_67_cast_fp16 = add(x = inputs_65_cast_fp16, y = x_41_cast_fp16)[name = tensor<string, []>("inputs_67_cast_fp16")];
             tensor<int32, [1]> out_67_axes_0 = const()[name = tensor<string, []>("out_67_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_2055_to_fp16 = const()[name = tensor<string, []>("op_2055_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_67_cast_fp16 = layer_norm(axes = out_67_axes_0, epsilon = var_2055_to_fp16, x = inputs_67_cast_fp16)[name = tensor<string, []>("out_67_cast_fp16")];
-            tensor<fp16, [1024]> input_191_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_191_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(345312896)))];
-            tensor<fp16, [1024]> input_191_beta_0_to_fp16 = const()[name = tensor<string, []>("input_191_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(345315008)))];
-            tensor<fp16, []> input_191_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_191_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_191_cast_fp16 = batch_norm(beta = input_191_beta_0_to_fp16, epsilon = input_191_epsilon_0_to_fp16, gamma = input_191_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_67_cast_fp16)[name = tensor<string, []>("input_191_cast_fp16")];
-            tensor<string, []> input_193_pad_type_0 = const()[name = tensor<string, []>("input_193_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_193_strides_0 = const()[name = tensor<string, []>("input_193_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_193_pad_0 = const()[name = tensor<string, []>("input_193_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_193_dilations_0 = const()[name = tensor<string, []>("input_193_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_193_groups_0 = const()[name = tensor<string, []>("input_193_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_2053_to_fp16 = const()[name = tensor<string, []>("op_2053_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_67_cast_fp16 = layer_norm(axes = out_67_axes_0, epsilon = var_2053_to_fp16, x = inputs_67_cast_fp16)[name = tensor<string, []>("out_67_cast_fp16")];
+            tensor<fp16, [1024]> input_193_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_193_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(345312896)))];
+            tensor<fp16, [1024]> input_193_beta_0_to_fp16 = const()[name = tensor<string, []>("input_193_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(345315008)))];
+            tensor<fp16, []> input_193_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_193_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_193_cast_fp16 = batch_norm(beta = input_193_beta_0_to_fp16, epsilon = input_193_epsilon_0_to_fp16, gamma = input_193_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_67_cast_fp16)[name = tensor<string, []>("input_193_cast_fp16")];
+            tensor<string, []> input_195_pad_type_0 = const()[name = tensor<string, []>("input_195_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_195_strides_0 = const()[name = tensor<string, []>("input_195_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_195_pad_0 = const()[name = tensor<string, []>("input_195_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_195_dilations_0 = const()[name = tensor<string, []>("input_195_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_195_groups_0 = const()[name = tensor<string, []>("input_195_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_6_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_feed_forward2_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(345317120)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_193_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_193_dilations_0, groups = input_193_groups_0, pad = input_193_pad_0, pad_type = input_193_pad_type_0, strides = input_193_strides_0, weight = layers_6_feed_forward2_fc1_weight_to_fp16, x = input_191_cast_fp16)[name = tensor<string, []>("input_193_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_195_cast_fp16 = silu(x = input_193_cast_fp16)[name = tensor<string, []>("input_195_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_195_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_195_dilations_0, groups = input_195_groups_0, pad = input_195_pad_0, pad_type = input_195_pad_type_0, strides = input_195_strides_0, weight = layers_6_feed_forward2_fc1_weight_to_fp16, x = input_193_cast_fp16)[name = tensor<string, []>("input_195_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_197_cast_fp16 = silu(x = input_195_cast_fp16)[name = tensor<string, []>("input_197_cast_fp16")];
             tensor<string, []> x_43_pad_type_0 = const()[name = tensor<string, []>("x_43_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_43_strides_0 = const()[name = tensor<string, []>("x_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_43_pad_0 = const()[name = tensor<string, []>("x_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_43_dilations_0 = const()[name = tensor<string, []>("x_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_43_groups_0 = const()[name = tensor<string, []>("x_43_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_2083_weight_0_to_fp16 = const()[name = tensor<string, []>("op_2083_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(353705792)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_2083_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_43_dilations_0, groups = x_43_groups_0, pad = x_43_pad_0, pad_type = x_43_pad_type_0, strides = x_43_strides_0, weight = var_2083_weight_0_to_fp16, x = input_195_cast_fp16)[name = tensor<string, []>("op_2083_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_69_cast_fp16 = add(x = inputs_67_cast_fp16, y = var_2083_cast_fp16)[name = tensor<string, []>("inputs_69_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_2081_weight_0_to_fp16 = const()[name = tensor<string, []>("op_2081_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(353705792)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_2081_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_43_dilations_0, groups = x_43_groups_0, pad = x_43_pad_0, pad_type = x_43_pad_type_0, strides = x_43_strides_0, weight = var_2081_weight_0_to_fp16, x = input_197_cast_fp16)[name = tensor<string, []>("op_2081_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_69_cast_fp16 = add(x = inputs_67_cast_fp16, y = var_2081_cast_fp16)[name = tensor<string, []>("inputs_69_cast_fp16")];
             tensor<int32, [1]> out_69_axes_0 = const()[name = tensor<string, []>("out_69_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_2093_to_fp16 = const()[name = tensor<string, []>("op_2093_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_69_cast_fp16 = layer_norm(axes = out_69_axes_0, epsilon = var_2093_to_fp16, x = inputs_69_cast_fp16)[name = tensor<string, []>("out_69_cast_fp16")];
+            tensor<fp16, []> var_2091_to_fp16 = const()[name = tensor<string, []>("op_2091_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_69_cast_fp16 = layer_norm(axes = out_69_axes_0, epsilon = var_2091_to_fp16, x = inputs_69_cast_fp16)[name = tensor<string, []>("out_69_cast_fp16")];
             tensor<fp16, [1024]> inputs_71_gamma_0_to_fp16 = const()[name = tensor<string, []>("inputs_71_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(362094464)))];
             tensor<fp16, [1024]> inputs_71_beta_0_to_fp16 = const()[name = tensor<string, []>("inputs_71_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(362096576)))];
             tensor<fp16, []> inputs_71_epsilon_0_to_fp16 = const()[name = tensor<string, []>("inputs_71_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_71_cast_fp16 = batch_norm(beta = inputs_71_beta_0_to_fp16, epsilon = inputs_71_epsilon_0_to_fp16, gamma = inputs_71_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_69_cast_fp16)[name = tensor<string, []>("inputs_71_cast_fp16")];
-            tensor<int32, []> var_2107 = const()[name = tensor<string, []>("op_2107"), val = tensor<int32, []>(3)];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_71_cast_fp16 = batch_norm(beta = inputs_71_beta_0_to_fp16, epsilon = inputs_71_epsilon_0_to_fp16, gamma = inputs_71_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_69_cast_fp16)[name = tensor<string, []>("inputs_71_cast_fp16")];
+            tensor<int32, []> var_2105 = const()[name = tensor<string, []>("op_2105"), val = tensor<int32, []>(3)];
             tensor<int32, [1]> out_71_axes_0 = const()[name = tensor<string, []>("out_71_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_2138_to_fp16 = const()[name = tensor<string, []>("op_2138_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_71_cast_fp16 = layer_norm(axes = out_71_axes_0, epsilon = var_2138_to_fp16, x = inputs_71_cast_fp16)[name = tensor<string, []>("out_71_cast_fp16")];
-            tensor<fp16, [1024]> input_197_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_197_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(362098688)))];
-            tensor<fp16, [1024]> input_197_beta_0_to_fp16 = const()[name = tensor<string, []>("input_197_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(362100800)))];
-            tensor<fp16, []> input_197_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_197_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_197_cast_fp16 = batch_norm(beta = input_197_beta_0_to_fp16, epsilon = input_197_epsilon_0_to_fp16, gamma = input_197_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_71_cast_fp16)[name = tensor<string, []>("input_197_cast_fp16")];
-            tensor<string, []> input_199_pad_type_0 = const()[name = tensor<string, []>("input_199_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_199_strides_0 = const()[name = tensor<string, []>("input_199_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_199_pad_0 = const()[name = tensor<string, []>("input_199_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_199_dilations_0 = const()[name = tensor<string, []>("input_199_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_199_groups_0 = const()[name = tensor<string, []>("input_199_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_2136_to_fp16 = const()[name = tensor<string, []>("op_2136_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_71_cast_fp16 = layer_norm(axes = out_71_axes_0, epsilon = var_2136_to_fp16, x = inputs_71_cast_fp16)[name = tensor<string, []>("out_71_cast_fp16")];
+            tensor<fp16, [1024]> input_199_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_199_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(362098688)))];
+            tensor<fp16, [1024]> input_199_beta_0_to_fp16 = const()[name = tensor<string, []>("input_199_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(362100800)))];
+            tensor<fp16, []> input_199_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_199_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_199_cast_fp16 = batch_norm(beta = input_199_beta_0_to_fp16, epsilon = input_199_epsilon_0_to_fp16, gamma = input_199_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_71_cast_fp16)[name = tensor<string, []>("input_199_cast_fp16")];
+            tensor<string, []> input_201_pad_type_0 = const()[name = tensor<string, []>("input_201_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_201_strides_0 = const()[name = tensor<string, []>("input_201_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_201_pad_0 = const()[name = tensor<string, []>("input_201_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_201_dilations_0 = const()[name = tensor<string, []>("input_201_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_201_groups_0 = const()[name = tensor<string, []>("input_201_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_7_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_feed_forward1_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(362102912)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_199_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_199_dilations_0, groups = input_199_groups_0, pad = input_199_pad_0, pad_type = input_199_pad_type_0, strides = input_199_strides_0, weight = layers_7_feed_forward1_fc1_weight_to_fp16, x = input_197_cast_fp16)[name = tensor<string, []>("input_199_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_201_cast_fp16 = silu(x = input_199_cast_fp16)[name = tensor<string, []>("input_201_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_201_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_201_dilations_0, groups = input_201_groups_0, pad = input_201_pad_0, pad_type = input_201_pad_type_0, strides = input_201_strides_0, weight = layers_7_feed_forward1_fc1_weight_to_fp16, x = input_199_cast_fp16)[name = tensor<string, []>("input_201_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_203_cast_fp16 = silu(x = input_201_cast_fp16)[name = tensor<string, []>("input_203_cast_fp16")];
             tensor<string, []> x_45_pad_type_0 = const()[name = tensor<string, []>("x_45_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_45_strides_0 = const()[name = tensor<string, []>("x_45_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_45_pad_0 = const()[name = tensor<string, []>("x_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_45_dilations_0 = const()[name = tensor<string, []>("x_45_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_45_groups_0 = const()[name = tensor<string, []>("x_45_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_2166_weight_0_to_fp16 = const()[name = tensor<string, []>("op_2166_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(370491584)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_2166_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_45_dilations_0, groups = x_45_groups_0, pad = x_45_pad_0, pad_type = x_45_pad_type_0, strides = x_45_strides_0, weight = var_2166_weight_0_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("op_2166_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_73_cast_fp16 = add(x = inputs_71_cast_fp16, y = var_2166_cast_fp16)[name = tensor<string, []>("inputs_73_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_2164_weight_0_to_fp16 = const()[name = tensor<string, []>("op_2164_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(370491584)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_2164_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_45_dilations_0, groups = x_45_groups_0, pad = x_45_pad_0, pad_type = x_45_pad_type_0, strides = x_45_strides_0, weight = var_2164_weight_0_to_fp16, x = input_203_cast_fp16)[name = tensor<string, []>("op_2164_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_73_cast_fp16 = add(x = inputs_71_cast_fp16, y = var_2164_cast_fp16)[name = tensor<string, []>("inputs_73_cast_fp16")];
             tensor<int32, [1]> out_73_axes_0 = const()[name = tensor<string, []>("out_73_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_2176_to_fp16 = const()[name = tensor<string, []>("op_2176_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_73_cast_fp16 = layer_norm(axes = out_73_axes_0, epsilon = var_2176_to_fp16, x = inputs_73_cast_fp16)[name = tensor<string, []>("out_73_cast_fp16")];
+            tensor<fp16, []> var_2174_to_fp16 = const()[name = tensor<string, []>("op_2174_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_73_cast_fp16 = layer_norm(axes = out_73_axes_0, epsilon = var_2174_to_fp16, x = inputs_73_cast_fp16)[name = tensor<string, []>("out_73_cast_fp16")];
             tensor<fp16, [1024]> obj_31_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_31_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(378880256)))];
             tensor<fp16, [1024]> obj_31_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_31_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(378882368)))];
             tensor<fp16, []> obj_31_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_31_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> obj_31_cast_fp16 = batch_norm(beta = obj_31_beta_0_to_fp16, epsilon = obj_31_epsilon_0_to_fp16, gamma = obj_31_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_73_cast_fp16)[name = tensor<string, []>("obj_31_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_31_cast_fp16 = batch_norm(beta = obj_31_beta_0_to_fp16, epsilon = obj_31_epsilon_0_to_fp16, gamma = obj_31_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_73_cast_fp16)[name = tensor<string, []>("obj_31_cast_fp16")];
             tensor<string, []> query_29_pad_type_0 = const()[name = tensor<string, []>("query_29_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> query_29_strides_0 = const()[name = tensor<string, []>("query_29_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> query_29_pad_0 = const()[name = tensor<string, []>("query_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> query_29_dilations_0 = const()[name = tensor<string, []>("query_29_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> query_29_groups_0 = const()[name = tensor<string, []>("query_29_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_7_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(378884480)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_29_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = query_29_dilations_0, groups = query_29_groups_0, pad = query_29_pad_0, pad_type = query_29_pad_type_0, strides = query_29_strides_0, weight = layers_7_self_attn_q_proj_weight_to_fp16, x = obj_31_cast_fp16)[name = tensor<string, []>("query_29_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> query_29_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_29_dilations_0, groups = query_29_groups_0, pad = query_29_pad_0, pad_type = query_29_pad_type_0, strides = query_29_strides_0, weight = layers_7_self_attn_q_proj_weight_to_fp16, x = obj_31_cast_fp16)[name = tensor<string, []>("query_29_cast_fp16")];
             tensor<string, []> key_15_pad_type_0 = const()[name = tensor<string, []>("key_15_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> key_15_strides_0 = const()[name = tensor<string, []>("key_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> key_15_pad_0 = const()[name = tensor<string, []>("key_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
@@ -1372,173 +1374,173 @@ program(1.0)
             tensor<int32, [2]> value_15_dilations_0 = const()[name = tensor<string, []>("value_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> value_15_groups_0 = const()[name = tensor<string, []>("value_15_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_7_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(383078912)))];
-            tensor<fp16, [1, 1024, 1, 188]> value_15_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = value_15_dilations_0, groups = value_15_groups_0, pad = value_15_pad_0, pad_type = value_15_pad_type_0, strides = value_15_strides_0, weight = layers_7_self_attn_v_proj_weight_to_fp16, x = obj_31_cast_fp16)[name = tensor<string, []>("value_15_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_2214_to_fp16 = const()[name = tensor<string, []>("op_2214_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(385176128)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_31_cast_fp16 = add(x = query_29_cast_fp16, y = var_2214_to_fp16)[name = tensor<string, []>("query_31_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_2217_to_fp16 = const()[name = tensor<string, []>("op_2217_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(385178240)))];
-            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_15_cast_fp16 = add(x = query_29_cast_fp16, y = var_2217_to_fp16)[name = tensor<string, []>("q_with_bias_v_15_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> value_15_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_15_dilations_0, groups = value_15_groups_0, pad = value_15_pad_0, pad_type = value_15_pad_type_0, strides = value_15_strides_0, weight = layers_7_self_attn_v_proj_weight_to_fp16, x = obj_31_cast_fp16)[name = tensor<string, []>("value_15_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_2212_to_fp16 = const()[name = tensor<string, []>("op_2212_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(385176128)))];
+            tensor<fp16, [1, 1024, 1, 188]> query_31_cast_fp16 = add(x = query_29_cast_fp16, y = var_2212_to_fp16)[name = tensor<string, []>("query_31_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_2215_to_fp16 = const()[name = tensor<string, []>("op_2215_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(385178240)))];
+            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_15_cast_fp16 = add(x = query_29_cast_fp16, y = var_2215_to_fp16)[name = tensor<string, []>("q_with_bias_v_15_cast_fp16")];
             tensor<string, []> p_15_pad_type_0 = const()[name = tensor<string, []>("p_15_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> p_15_strides_0 = const()[name = tensor<string, []>("p_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> p_15_pad_0 = const()[name = tensor<string, []>("p_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> p_15_dilations_0 = const()[name = tensor<string, []>("p_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> p_15_groups_0 = const()[name = tensor<string, []>("p_15_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_7_self_attn_linear_pos_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_linear_pos_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(385180352)))];
-            tensor<fp16, [1, 1024, 1, 375]> p_15_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = p_15_dilations_0, groups = p_15_groups_0, pad = p_15_pad_0, pad_type = p_15_pad_type_0, strides = p_15_strides_0, weight = layers_7_self_attn_linear_pos_weight_to_fp16, x = pos_enc_to_fp16)[name = tensor<string, []>("p_15_cast_fp16")];
-            tensor<int32, [4]> var_2228 = const()[name = tensor<string, []>("op_2228"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_2229_cast_fp16 = reshape(shape = var_2228, x = q_with_bias_v_15_cast_fp16)[name = tensor<string, []>("op_2229_cast_fp16")];
-            tensor<int32, [4]> var_2230 = const()[name = tensor<string, []>("op_2230"), val = tensor<int32, [4]>([1, 8, 128, -1])];
-            tensor<fp16, [1, 8, 128, 375]> var_2231_cast_fp16 = reshape(shape = var_2230, x = p_15_cast_fp16)[name = tensor<string, []>("op_2231_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 375]> p_15_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_15_dilations_0, groups = p_15_groups_0, pad = p_15_pad_0, pad_type = p_15_pad_type_0, strides = p_15_strides_0, weight = layers_7_self_attn_linear_pos_weight_to_fp16, x = obj_3_cast_fp16)[name = tensor<string, []>("p_15_cast_fp16")];
+            tensor<int32, [4]> var_2226 = const()[name = tensor<string, []>("op_2226"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_2227_cast_fp16 = reshape(shape = var_2226, x = q_with_bias_v_15_cast_fp16)[name = tensor<string, []>("op_2227_cast_fp16")];
+            tensor<int32, [4]> var_2228 = const()[name = tensor<string, []>("op_2228"), val = tensor<int32, [4]>([1, 8, 128, -1])];
+            tensor<fp16, [1, 8, 128, 375]> var_2229_cast_fp16 = reshape(shape = var_2228, x = p_15_cast_fp16)[name = tensor<string, []>("op_2229_cast_fp16")];
             tensor<bool, []> matrix_bd_57_transpose_x_0 = const()[name = tensor<string, []>("matrix_bd_57_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> matrix_bd_57_transpose_y_0 = const()[name = tensor<string, []>("matrix_bd_57_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_57_cast_fp16 = matmul(transpose_x = matrix_bd_57_transpose_x_0, transpose_y = matrix_bd_57_transpose_y_0, x = var_2229_cast_fp16, y = var_2231_cast_fp16)[name = tensor<string, []>("matrix_bd_57_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_57_cast_fp16 = matmul(transpose_x = matrix_bd_57_transpose_x_0, transpose_y = matrix_bd_57_transpose_y_0, x = var_2227_cast_fp16, y = var_2229_cast_fp16)[name = tensor<string, []>("matrix_bd_57_cast_fp16")];
             tensor<int32, [8]> matrix_bd_59_pad_0 = const()[name = tensor<string, []>("matrix_bd_59_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
             tensor<string, []> matrix_bd_59_mode_0 = const()[name = tensor<string, []>("matrix_bd_59_mode_0"), val = tensor<string, []>("constant")];
             tensor<fp16, []> const_87_to_fp16 = const()[name = tensor<string, []>("const_87_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
             tensor<fp16, [1, 8, 188, 376]> matrix_bd_59_cast_fp16 = pad(constant_val = const_87_to_fp16, mode = matrix_bd_59_mode_0, pad = matrix_bd_59_pad_0, x = matrix_bd_57_cast_fp16)[name = tensor<string, []>("matrix_bd_59_cast_fp16")];
-            tensor<int32, [4]> var_2240 = const()[name = tensor<string, []>("op_2240"), val = tensor<int32, [4]>([1, 8, -1, 188])];
-            tensor<fp16, [1, 8, 376, 188]> matrix_bd_61_cast_fp16 = reshape(shape = var_2240, x = matrix_bd_59_cast_fp16)[name = tensor<string, []>("matrix_bd_61_cast_fp16")];
-            tensor<int32, [4]> var_2244_begin_0 = const()[name = tensor<string, []>("op_2244_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
-            tensor<int32, [4]> var_2244_end_0 = const()[name = tensor<string, []>("op_2244_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
-            tensor<bool, [4]> var_2244_end_mask_0 = const()[name = tensor<string, []>("op_2244_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
-            tensor<fp16, [1, 8, 375, 188]> var_2244_cast_fp16 = slice_by_index(begin = var_2244_begin_0, end = var_2244_end_0, end_mask = var_2244_end_mask_0, x = matrix_bd_61_cast_fp16)[name = tensor<string, []>("op_2244_cast_fp16")];
-            tensor<int32, [4]> var_2245 = const()[name = tensor<string, []>("op_2245"), val = tensor<int32, [4]>([1, 8, 188, 375])];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_63_cast_fp16 = reshape(shape = var_2245, x = var_2244_cast_fp16)[name = tensor<string, []>("matrix_bd_63_cast_fp16")];
-            tensor<int32, [4]> var_2250_begin_0 = const()[name = tensor<string, []>("op_2250_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [4]> var_2250_end_0 = const()[name = tensor<string, []>("op_2250_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
-            tensor<bool, [4]> var_2250_end_mask_0 = const()[name = tensor<string, []>("op_2250_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
-            tensor<fp16, [1, 8, 188, 188]> var_2250_cast_fp16 = slice_by_index(begin = var_2250_begin_0, end = var_2250_end_0, end_mask = var_2250_end_mask_0, x = matrix_bd_63_cast_fp16)[name = tensor<string, []>("op_2250_cast_fp16")];
-            tensor<fp16, []> var_2251_to_fp16 = const()[name = tensor<string, []>("op_2251_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 188, 188]> qk_mask_15_cast_fp16 = mul(x = var_2250_cast_fp16, y = var_2251_to_fp16)[name = tensor<string, []>("qk_mask_15_cast_fp16")];
-            tensor<int32, [4]> var_2255 = const()[name = tensor<string, []>("op_2255"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> mh_q_15_cast_fp16 = reshape(shape = var_2255, x = query_31_cast_fp16)[name = tensor<string, []>("mh_q_15_cast_fp16")];
-            tensor<fp16, []> var_2257_to_fp16 = const()[name = tensor<string, []>("op_2257_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 128, 188]> var_2258_cast_fp16 = mul(x = mh_q_15_cast_fp16, y = var_2257_to_fp16)[name = tensor<string, []>("op_2258_cast_fp16")];
-            tensor<int32, [4]> var_2261 = const()[name = tensor<string, []>("op_2261"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_2262_cast_fp16 = reshape(shape = var_2261, x = key_15_cast_fp16)[name = tensor<string, []>("op_2262_cast_fp16")];
+            tensor<int32, [4]> var_2238 = const()[name = tensor<string, []>("op_2238"), val = tensor<int32, [4]>([1, 8, -1, 188])];
+            tensor<fp16, [1, 8, 376, 188]> matrix_bd_61_cast_fp16 = reshape(shape = var_2238, x = matrix_bd_59_cast_fp16)[name = tensor<string, []>("matrix_bd_61_cast_fp16")];
+            tensor<int32, [4]> var_2242_begin_0 = const()[name = tensor<string, []>("op_2242_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2242_end_0 = const()[name = tensor<string, []>("op_2242_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
+            tensor<bool, [4]> var_2242_end_mask_0 = const()[name = tensor<string, []>("op_2242_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 375, 188]> var_2242_cast_fp16 = slice_by_index(begin = var_2242_begin_0, end = var_2242_end_0, end_mask = var_2242_end_mask_0, x = matrix_bd_61_cast_fp16)[name = tensor<string, []>("op_2242_cast_fp16")];
+            tensor<int32, [4]> var_2243 = const()[name = tensor<string, []>("op_2243"), val = tensor<int32, [4]>([1, 8, 188, 375])];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_63_cast_fp16 = reshape(shape = var_2243, x = var_2242_cast_fp16)[name = tensor<string, []>("matrix_bd_63_cast_fp16")];
+            tensor<int32, [4]> var_2248_begin_0 = const()[name = tensor<string, []>("op_2248_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_2248_end_0 = const()[name = tensor<string, []>("op_2248_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
+            tensor<bool, [4]> var_2248_end_mask_0 = const()[name = tensor<string, []>("op_2248_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 188, 188]> var_2248_cast_fp16 = slice_by_index(begin = var_2248_begin_0, end = var_2248_end_0, end_mask = var_2248_end_mask_0, x = matrix_bd_63_cast_fp16)[name = tensor<string, []>("op_2248_cast_fp16")];
+            tensor<fp16, []> var_2249_to_fp16 = const()[name = tensor<string, []>("op_2249_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 188, 188]> qk_mask_15_cast_fp16 = mul(x = var_2248_cast_fp16, y = var_2249_to_fp16)[name = tensor<string, []>("qk_mask_15_cast_fp16")];
+            tensor<int32, [4]> var_2253 = const()[name = tensor<string, []>("op_2253"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> mh_q_15_cast_fp16 = reshape(shape = var_2253, x = query_31_cast_fp16)[name = tensor<string, []>("mh_q_15_cast_fp16")];
+            tensor<fp16, []> var_2255_to_fp16 = const()[name = tensor<string, []>("op_2255_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 128, 188]> var_2256_cast_fp16 = mul(x = mh_q_15_cast_fp16, y = var_2255_to_fp16)[name = tensor<string, []>("op_2256_cast_fp16")];
+            tensor<int32, [4]> var_2259 = const()[name = tensor<string, []>("op_2259"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_2260_cast_fp16 = reshape(shape = var_2259, x = key_15_cast_fp16)[name = tensor<string, []>("op_2260_cast_fp16")];
             tensor<bool, []> mh_w_29_transpose_x_0 = const()[name = tensor<string, []>("mh_w_29_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> mh_w_29_transpose_y_0 = const()[name = tensor<string, []>("mh_w_29_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 188]> mh_w_29_cast_fp16 = matmul(transpose_x = mh_w_29_transpose_x_0, transpose_y = mh_w_29_transpose_y_0, x = var_2258_cast_fp16, y = var_2262_cast_fp16)[name = tensor<string, []>("mh_w_29_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> mh_w_29_cast_fp16 = matmul(transpose_x = mh_w_29_transpose_x_0, transpose_y = mh_w_29_transpose_y_0, x = var_2256_cast_fp16, y = var_2260_cast_fp16)[name = tensor<string, []>("mh_w_29_cast_fp16")];
             tensor<fp16, [1, 8, 188, 188]> mh_w_31_cast_fp16 = add(x = mh_w_29_cast_fp16, y = qk_mask_15_cast_fp16)[name = tensor<string, []>("mh_w_31_cast_fp16")];
-            tensor<fp16, [1, 8, 188, 188]> var_2266_cast_fp16 = softmax(axis = var_2107, x = mh_w_31_cast_fp16)[name = tensor<string, []>("op_2266_cast_fp16")];
-            tensor<int32, [4]> var_2267 = const()[name = tensor<string, []>("op_2267"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_2268_cast_fp16 = reshape(shape = var_2267, x = value_15_cast_fp16)[name = tensor<string, []>("op_2268_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> var_2264_cast_fp16 = softmax(axis = var_2105, x = mh_w_31_cast_fp16)[name = tensor<string, []>("op_2264_cast_fp16")];
+            tensor<int32, [4]> var_2265 = const()[name = tensor<string, []>("op_2265"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_2266_cast_fp16 = reshape(shape = var_2265, x = value_15_cast_fp16)[name = tensor<string, []>("op_2266_cast_fp16")];
             tensor<bool, []> attn_15_transpose_x_0 = const()[name = tensor<string, []>("attn_15_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> attn_15_transpose_y_0 = const()[name = tensor<string, []>("attn_15_transpose_y_0"), val = tensor<bool, []>(true)];
-            tensor<fp16, [1, 8, 128, 188]> attn_15_cast_fp16 = matmul(transpose_x = attn_15_transpose_x_0, transpose_y = attn_15_transpose_y_0, x = var_2268_cast_fp16, y = var_2266_cast_fp16)[name = tensor<string, []>("attn_15_cast_fp16")];
-            tensor<int32, [4]> var_2271 = const()[name = tensor<string, []>("op_2271"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
-            tensor<fp16, [1, 1024, 1, 188]> input_203_cast_fp16 = reshape(shape = var_2271, x = attn_15_cast_fp16)[name = tensor<string, []>("input_203_cast_fp16")];
+            tensor<fp16, [1, 8, 128, 188]> attn_15_cast_fp16 = matmul(transpose_x = attn_15_transpose_x_0, transpose_y = attn_15_transpose_y_0, x = var_2266_cast_fp16, y = var_2264_cast_fp16)[name = tensor<string, []>("attn_15_cast_fp16")];
+            tensor<int32, [4]> var_2269 = const()[name = tensor<string, []>("op_2269"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
+            tensor<fp16, [1, 1024, 1, 188]> input_205_cast_fp16 = reshape(shape = var_2269, x = attn_15_cast_fp16)[name = tensor<string, []>("input_205_cast_fp16")];
             tensor<string, []> obj_33_pad_type_0 = const()[name = tensor<string, []>("obj_33_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> obj_33_strides_0 = const()[name = tensor<string, []>("obj_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> obj_33_pad_0 = const()[name = tensor<string, []>("obj_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> obj_33_dilations_0 = const()[name = tensor<string, []>("obj_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> obj_33_groups_0 = const()[name = tensor<string, []>("obj_33_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_7_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(387277568)))];
-            tensor<fp16, [1, 1024, 1, 188]> obj_33_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = obj_33_dilations_0, groups = obj_33_groups_0, pad = obj_33_pad_0, pad_type = obj_33_pad_type_0, strides = obj_33_strides_0, weight = layers_7_self_attn_o_proj_weight_to_fp16, x = input_203_cast_fp16)[name = tensor<string, []>("obj_33_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_33_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_33_dilations_0, groups = obj_33_groups_0, pad = obj_33_pad_0, pad_type = obj_33_pad_type_0, strides = obj_33_strides_0, weight = layers_7_self_attn_o_proj_weight_to_fp16, x = input_205_cast_fp16)[name = tensor<string, []>("obj_33_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_75_cast_fp16 = add(x = inputs_73_cast_fp16, y = obj_33_cast_fp16)[name = tensor<string, []>("inputs_75_cast_fp16")];
             tensor<int32, [1]> out_75_axes_0 = const()[name = tensor<string, []>("out_75_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_2289_to_fp16 = const()[name = tensor<string, []>("op_2289_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_75_cast_fp16 = layer_norm(axes = out_75_axes_0, epsilon = var_2289_to_fp16, x = inputs_75_cast_fp16)[name = tensor<string, []>("out_75_cast_fp16")];
-            tensor<fp16, [1024]> input_205_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_205_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(389374784)))];
-            tensor<fp16, [1024]> input_205_beta_0_to_fp16 = const()[name = tensor<string, []>("input_205_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(389376896)))];
-            tensor<fp16, []> input_205_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_205_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_205_cast_fp16 = batch_norm(beta = input_205_beta_0_to_fp16, epsilon = input_205_epsilon_0_to_fp16, gamma = input_205_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_75_cast_fp16)[name = tensor<string, []>("input_205_cast_fp16")];
-            tensor<string, []> input_207_pad_type_0 = const()[name = tensor<string, []>("input_207_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_207_strides_0 = const()[name = tensor<string, []>("input_207_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_207_pad_0 = const()[name = tensor<string, []>("input_207_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_207_dilations_0 = const()[name = tensor<string, []>("input_207_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_207_groups_0 = const()[name = tensor<string, []>("input_207_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_2287_to_fp16 = const()[name = tensor<string, []>("op_2287_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_75_cast_fp16 = layer_norm(axes = out_75_axes_0, epsilon = var_2287_to_fp16, x = inputs_75_cast_fp16)[name = tensor<string, []>("out_75_cast_fp16")];
+            tensor<fp16, [1024]> input_207_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_207_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(389374784)))];
+            tensor<fp16, [1024]> input_207_beta_0_to_fp16 = const()[name = tensor<string, []>("input_207_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(389376896)))];
+            tensor<fp16, []> input_207_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_207_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_207_cast_fp16 = batch_norm(beta = input_207_beta_0_to_fp16, epsilon = input_207_epsilon_0_to_fp16, gamma = input_207_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_75_cast_fp16)[name = tensor<string, []>("input_207_cast_fp16")];
+            tensor<string, []> input_209_pad_type_0 = const()[name = tensor<string, []>("input_209_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_209_strides_0 = const()[name = tensor<string, []>("input_209_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_209_pad_0 = const()[name = tensor<string, []>("input_209_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_209_dilations_0 = const()[name = tensor<string, []>("input_209_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_209_groups_0 = const()[name = tensor<string, []>("input_209_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [2048, 1024, 1, 1]> layers_7_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(389379008)))];
-            tensor<fp16, [1, 2048, 1, 188]> input_207_cast_fp16 = conv(dilations = input_207_dilations_0, groups = input_207_groups_0, pad = input_207_pad_0, pad_type = input_207_pad_type_0, strides = input_207_strides_0, weight = layers_7_conv_pointwise_conv1_weight_to_fp16, x = input_205_cast_fp16)[name = tensor<string, []>("input_207_cast_fp16")];
-            tensor<int32, []> input_209_split_num_splits_0 = const()[name = tensor<string, []>("input_209_split_num_splits_0"), val = tensor<int32, []>(2)];
-            tensor<int32, []> input_209_split_axis_0 = const()[name = tensor<string, []>("input_209_split_axis_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1, 1024, 1, 188]> input_209_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_209_split_cast_fp16_1 = split(axis = input_209_split_axis_0, num_splits = input_209_split_num_splits_0, x = input_207_cast_fp16)[name = tensor<string, []>("input_209_split_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_209_split_1_sigmoid_cast_fp16 = sigmoid(x = input_209_split_cast_fp16_1)[name = tensor<string, []>("input_209_split_1_sigmoid_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_209_cast_fp16 = mul(x = input_209_split_cast_fp16_0, y = input_209_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_209_cast_fp16")];
-            tensor<string, []> input_211_pad_type_0 = const()[name = tensor<string, []>("input_211_pad_type_0"), val = tensor<string, []>("custom")];
-            tensor<int32, [4]> input_211_pad_0 = const()[name = tensor<string, []>("input_211_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
-            tensor<int32, []> input_211_groups_0 = const()[name = tensor<string, []>("input_211_groups_0"), val = tensor<int32, []>(1024)];
-            tensor<int32, [2]> input_211_strides_0 = const()[name = tensor<string, []>("input_211_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> input_211_dilations_0 = const()[name = tensor<string, []>("input_211_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [1, 2048, 1, 188]> input_209_cast_fp16 = conv(dilations = input_209_dilations_0, groups = input_209_groups_0, pad = input_209_pad_0, pad_type = input_209_pad_type_0, strides = input_209_strides_0, weight = layers_7_conv_pointwise_conv1_weight_to_fp16, x = input_207_cast_fp16)[name = tensor<string, []>("input_209_cast_fp16")];
+            tensor<int32, []> input_211_split_num_splits_0 = const()[name = tensor<string, []>("input_211_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> input_211_split_axis_0 = const()[name = tensor<string, []>("input_211_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 1, 188]> input_211_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_211_split_cast_fp16_1 = split(axis = input_211_split_axis_0, num_splits = input_211_split_num_splits_0, x = input_209_cast_fp16)[name = tensor<string, []>("input_211_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_211_split_1_sigmoid_cast_fp16 = sigmoid(x = input_211_split_cast_fp16_1)[name = tensor<string, []>("input_211_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_211_cast_fp16 = mul(x = input_211_split_cast_fp16_0, y = input_211_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_211_cast_fp16")];
+            tensor<string, []> input_213_pad_type_0 = const()[name = tensor<string, []>("input_213_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_213_pad_0 = const()[name = tensor<string, []>("input_213_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
+            tensor<int32, []> input_213_groups_0 = const()[name = tensor<string, []>("input_213_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [2]> input_213_strides_0 = const()[name = tensor<string, []>("input_213_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> input_213_dilations_0 = const()[name = tensor<string, []>("input_213_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<fp16, [1024, 1, 1, 9]> const_282_to_fp16 = const()[name = tensor<string, []>("const_282_to_fp16"), val = tensor<fp16, [1024, 1, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(393573376)))];
             tensor<fp16, [1024]> const_283_to_fp16 = const()[name = tensor<string, []>("const_283_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(393591872)))];
-            tensor<fp16, [1, 1024, 1, 188]> input_213_cast_fp16 = conv(bias = const_283_to_fp16, dilations = input_211_dilations_0, groups = input_211_groups_0, pad = input_211_pad_0, pad_type = input_211_pad_type_0, strides = input_211_strides_0, weight = const_282_to_fp16, x = input_209_cast_fp16)[name = tensor<string, []>("input_213_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_215_cast_fp16 = silu(x = input_213_cast_fp16)[name = tensor<string, []>("input_215_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_215_cast_fp16 = conv(bias = const_283_to_fp16, dilations = input_213_dilations_0, groups = input_213_groups_0, pad = input_213_pad_0, pad_type = input_213_pad_type_0, strides = input_213_strides_0, weight = const_282_to_fp16, x = input_211_cast_fp16)[name = tensor<string, []>("input_215_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_217_cast_fp16 = silu(x = input_215_cast_fp16)[name = tensor<string, []>("input_217_cast_fp16")];
             tensor<string, []> x_47_pad_type_0 = const()[name = tensor<string, []>("x_47_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_47_strides_0 = const()[name = tensor<string, []>("x_47_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_47_pad_0 = const()[name = tensor<string, []>("x_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_47_dilations_0 = const()[name = tensor<string, []>("x_47_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_47_groups_0 = const()[name = tensor<string, []>("x_47_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_7_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(393593984)))];
-            tensor<fp16, [1, 1024, 1, 188]> x_47_cast_fp16 = conv(dilations = x_47_dilations_0, groups = x_47_groups_0, pad = x_47_pad_0, pad_type = x_47_pad_type_0, strides = x_47_strides_0, weight = layers_7_conv_pointwise_conv2_weight_to_fp16, x = input_215_cast_fp16)[name = tensor<string, []>("x_47_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> x_47_cast_fp16 = conv(dilations = x_47_dilations_0, groups = x_47_groups_0, pad = x_47_pad_0, pad_type = x_47_pad_type_0, strides = x_47_strides_0, weight = layers_7_conv_pointwise_conv2_weight_to_fp16, x = input_217_cast_fp16)[name = tensor<string, []>("x_47_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_77_cast_fp16 = add(x = inputs_75_cast_fp16, y = x_47_cast_fp16)[name = tensor<string, []>("inputs_77_cast_fp16")];
             tensor<int32, [1]> out_77_axes_0 = const()[name = tensor<string, []>("out_77_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_2337_to_fp16 = const()[name = tensor<string, []>("op_2337_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_77_cast_fp16 = layer_norm(axes = out_77_axes_0, epsilon = var_2337_to_fp16, x = inputs_77_cast_fp16)[name = tensor<string, []>("out_77_cast_fp16")];
-            tensor<fp16, [1024]> input_217_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_217_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(395691200)))];
-            tensor<fp16, [1024]> input_217_beta_0_to_fp16 = const()[name = tensor<string, []>("input_217_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(395693312)))];
-            tensor<fp16, []> input_217_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_217_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_217_cast_fp16 = batch_norm(beta = input_217_beta_0_to_fp16, epsilon = input_217_epsilon_0_to_fp16, gamma = input_217_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_77_cast_fp16)[name = tensor<string, []>("input_217_cast_fp16")];
-            tensor<string, []> input_219_pad_type_0 = const()[name = tensor<string, []>("input_219_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_219_strides_0 = const()[name = tensor<string, []>("input_219_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_219_pad_0 = const()[name = tensor<string, []>("input_219_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_219_dilations_0 = const()[name = tensor<string, []>("input_219_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_219_groups_0 = const()[name = tensor<string, []>("input_219_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_2335_to_fp16 = const()[name = tensor<string, []>("op_2335_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_77_cast_fp16 = layer_norm(axes = out_77_axes_0, epsilon = var_2335_to_fp16, x = inputs_77_cast_fp16)[name = tensor<string, []>("out_77_cast_fp16")];
+            tensor<fp16, [1024]> input_219_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_219_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(395691200)))];
+            tensor<fp16, [1024]> input_219_beta_0_to_fp16 = const()[name = tensor<string, []>("input_219_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(395693312)))];
+            tensor<fp16, []> input_219_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_219_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_219_cast_fp16 = batch_norm(beta = input_219_beta_0_to_fp16, epsilon = input_219_epsilon_0_to_fp16, gamma = input_219_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_77_cast_fp16)[name = tensor<string, []>("input_219_cast_fp16")];
+            tensor<string, []> input_221_pad_type_0 = const()[name = tensor<string, []>("input_221_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_221_strides_0 = const()[name = tensor<string, []>("input_221_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_221_pad_0 = const()[name = tensor<string, []>("input_221_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_221_dilations_0 = const()[name = tensor<string, []>("input_221_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_221_groups_0 = const()[name = tensor<string, []>("input_221_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_7_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_feed_forward2_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(395695424)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_219_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_219_dilations_0, groups = input_219_groups_0, pad = input_219_pad_0, pad_type = input_219_pad_type_0, strides = input_219_strides_0, weight = layers_7_feed_forward2_fc1_weight_to_fp16, x = input_217_cast_fp16)[name = tensor<string, []>("input_219_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_221_cast_fp16 = silu(x = input_219_cast_fp16)[name = tensor<string, []>("input_221_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_221_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_221_dilations_0, groups = input_221_groups_0, pad = input_221_pad_0, pad_type = input_221_pad_type_0, strides = input_221_strides_0, weight = layers_7_feed_forward2_fc1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor<string, []>("input_221_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_223_cast_fp16 = silu(x = input_221_cast_fp16)[name = tensor<string, []>("input_223_cast_fp16")];
             tensor<string, []> x_49_pad_type_0 = const()[name = tensor<string, []>("x_49_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_49_strides_0 = const()[name = tensor<string, []>("x_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_49_pad_0 = const()[name = tensor<string, []>("x_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_49_dilations_0 = const()[name = tensor<string, []>("x_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_49_groups_0 = const()[name = tensor<string, []>("x_49_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_2365_weight_0_to_fp16 = const()[name = tensor<string, []>("op_2365_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(404084096)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_2365_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_49_dilations_0, groups = x_49_groups_0, pad = x_49_pad_0, pad_type = x_49_pad_type_0, strides = x_49_strides_0, weight = var_2365_weight_0_to_fp16, x = input_221_cast_fp16)[name = tensor<string, []>("op_2365_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_79_cast_fp16 = add(x = inputs_77_cast_fp16, y = var_2365_cast_fp16)[name = tensor<string, []>("inputs_79_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_2363_weight_0_to_fp16 = const()[name = tensor<string, []>("op_2363_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(404084096)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_2363_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_49_dilations_0, groups = x_49_groups_0, pad = x_49_pad_0, pad_type = x_49_pad_type_0, strides = x_49_strides_0, weight = var_2363_weight_0_to_fp16, x = input_223_cast_fp16)[name = tensor<string, []>("op_2363_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_79_cast_fp16 = add(x = inputs_77_cast_fp16, y = var_2363_cast_fp16)[name = tensor<string, []>("inputs_79_cast_fp16")];
             tensor<int32, [1]> out_79_axes_0 = const()[name = tensor<string, []>("out_79_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_2375_to_fp16 = const()[name = tensor<string, []>("op_2375_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_79_cast_fp16 = layer_norm(axes = out_79_axes_0, epsilon = var_2375_to_fp16, x = inputs_79_cast_fp16)[name = tensor<string, []>("out_79_cast_fp16")];
+            tensor<fp16, []> var_2373_to_fp16 = const()[name = tensor<string, []>("op_2373_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_79_cast_fp16 = layer_norm(axes = out_79_axes_0, epsilon = var_2373_to_fp16, x = inputs_79_cast_fp16)[name = tensor<string, []>("out_79_cast_fp16")];
             tensor<fp16, [1024]> inputs_81_gamma_0_to_fp16 = const()[name = tensor<string, []>("inputs_81_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(412472768)))];
             tensor<fp16, [1024]> inputs_81_beta_0_to_fp16 = const()[name = tensor<string, []>("inputs_81_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(412474880)))];
             tensor<fp16, []> inputs_81_epsilon_0_to_fp16 = const()[name = tensor<string, []>("inputs_81_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_81_cast_fp16 = batch_norm(beta = inputs_81_beta_0_to_fp16, epsilon = inputs_81_epsilon_0_to_fp16, gamma = inputs_81_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_79_cast_fp16)[name = tensor<string, []>("inputs_81_cast_fp16")];
-            tensor<int32, []> var_2389 = const()[name = tensor<string, []>("op_2389"), val = tensor<int32, []>(3)];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_81_cast_fp16 = batch_norm(beta = inputs_81_beta_0_to_fp16, epsilon = inputs_81_epsilon_0_to_fp16, gamma = inputs_81_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_79_cast_fp16)[name = tensor<string, []>("inputs_81_cast_fp16")];
+            tensor<int32, []> var_2387 = const()[name = tensor<string, []>("op_2387"), val = tensor<int32, []>(3)];
             tensor<int32, [1]> out_81_axes_0 = const()[name = tensor<string, []>("out_81_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_2420_to_fp16 = const()[name = tensor<string, []>("op_2420_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_81_cast_fp16 = layer_norm(axes = out_81_axes_0, epsilon = var_2420_to_fp16, x = inputs_81_cast_fp16)[name = tensor<string, []>("out_81_cast_fp16")];
-            tensor<fp16, [1024]> input_223_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_223_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(412476992)))];
-            tensor<fp16, [1024]> input_223_beta_0_to_fp16 = const()[name = tensor<string, []>("input_223_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(412479104)))];
-            tensor<fp16, []> input_223_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_223_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_223_cast_fp16 = batch_norm(beta = input_223_beta_0_to_fp16, epsilon = input_223_epsilon_0_to_fp16, gamma = input_223_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_81_cast_fp16)[name = tensor<string, []>("input_223_cast_fp16")];
-            tensor<string, []> input_225_pad_type_0 = const()[name = tensor<string, []>("input_225_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_225_strides_0 = const()[name = tensor<string, []>("input_225_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_225_pad_0 = const()[name = tensor<string, []>("input_225_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_225_dilations_0 = const()[name = tensor<string, []>("input_225_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_225_groups_0 = const()[name = tensor<string, []>("input_225_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_2418_to_fp16 = const()[name = tensor<string, []>("op_2418_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_81_cast_fp16 = layer_norm(axes = out_81_axes_0, epsilon = var_2418_to_fp16, x = inputs_81_cast_fp16)[name = tensor<string, []>("out_81_cast_fp16")];
+            tensor<fp16, [1024]> input_225_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_225_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(412476992)))];
+            tensor<fp16, [1024]> input_225_beta_0_to_fp16 = const()[name = tensor<string, []>("input_225_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(412479104)))];
+            tensor<fp16, []> input_225_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_225_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_225_cast_fp16 = batch_norm(beta = input_225_beta_0_to_fp16, epsilon = input_225_epsilon_0_to_fp16, gamma = input_225_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_81_cast_fp16)[name = tensor<string, []>("input_225_cast_fp16")];
+            tensor<string, []> input_227_pad_type_0 = const()[name = tensor<string, []>("input_227_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_227_strides_0 = const()[name = tensor<string, []>("input_227_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_227_pad_0 = const()[name = tensor<string, []>("input_227_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_227_dilations_0 = const()[name = tensor<string, []>("input_227_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_227_groups_0 = const()[name = tensor<string, []>("input_227_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_8_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_feed_forward1_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(412481216)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_225_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_225_dilations_0, groups = input_225_groups_0, pad = input_225_pad_0, pad_type = input_225_pad_type_0, strides = input_225_strides_0, weight = layers_8_feed_forward1_fc1_weight_to_fp16, x = input_223_cast_fp16)[name = tensor<string, []>("input_225_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_227_cast_fp16 = silu(x = input_225_cast_fp16)[name = tensor<string, []>("input_227_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_227_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_227_dilations_0, groups = input_227_groups_0, pad = input_227_pad_0, pad_type = input_227_pad_type_0, strides = input_227_strides_0, weight = layers_8_feed_forward1_fc1_weight_to_fp16, x = input_225_cast_fp16)[name = tensor<string, []>("input_227_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_229_cast_fp16 = silu(x = input_227_cast_fp16)[name = tensor<string, []>("input_229_cast_fp16")];
             tensor<string, []> x_51_pad_type_0 = const()[name = tensor<string, []>("x_51_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_51_strides_0 = const()[name = tensor<string, []>("x_51_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_51_pad_0 = const()[name = tensor<string, []>("x_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_51_dilations_0 = const()[name = tensor<string, []>("x_51_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_51_groups_0 = const()[name = tensor<string, []>("x_51_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_2448_weight_0_to_fp16 = const()[name = tensor<string, []>("op_2448_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(420869888)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_2448_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_51_dilations_0, groups = x_51_groups_0, pad = x_51_pad_0, pad_type = x_51_pad_type_0, strides = x_51_strides_0, weight = var_2448_weight_0_to_fp16, x = input_227_cast_fp16)[name = tensor<string, []>("op_2448_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_83_cast_fp16 = add(x = inputs_81_cast_fp16, y = var_2448_cast_fp16)[name = tensor<string, []>("inputs_83_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_2446_weight_0_to_fp16 = const()[name = tensor<string, []>("op_2446_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(420869888)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_2446_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_51_dilations_0, groups = x_51_groups_0, pad = x_51_pad_0, pad_type = x_51_pad_type_0, strides = x_51_strides_0, weight = var_2446_weight_0_to_fp16, x = input_229_cast_fp16)[name = tensor<string, []>("op_2446_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_83_cast_fp16 = add(x = inputs_81_cast_fp16, y = var_2446_cast_fp16)[name = tensor<string, []>("inputs_83_cast_fp16")];
             tensor<int32, [1]> out_83_axes_0 = const()[name = tensor<string, []>("out_83_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_2458_to_fp16 = const()[name = tensor<string, []>("op_2458_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_83_cast_fp16 = layer_norm(axes = out_83_axes_0, epsilon = var_2458_to_fp16, x = inputs_83_cast_fp16)[name = tensor<string, []>("out_83_cast_fp16")];
+            tensor<fp16, []> var_2456_to_fp16 = const()[name = tensor<string, []>("op_2456_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_83_cast_fp16 = layer_norm(axes = out_83_axes_0, epsilon = var_2456_to_fp16, x = inputs_83_cast_fp16)[name = tensor<string, []>("out_83_cast_fp16")];
             tensor<fp16, [1024]> obj_35_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_35_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(429258560)))];
             tensor<fp16, [1024]> obj_35_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_35_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(429260672)))];
             tensor<fp16, []> obj_35_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_35_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> obj_35_cast_fp16 = batch_norm(beta = obj_35_beta_0_to_fp16, epsilon = obj_35_epsilon_0_to_fp16, gamma = obj_35_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_83_cast_fp16)[name = tensor<string, []>("obj_35_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_35_cast_fp16 = batch_norm(beta = obj_35_beta_0_to_fp16, epsilon = obj_35_epsilon_0_to_fp16, gamma = obj_35_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_83_cast_fp16)[name = tensor<string, []>("obj_35_cast_fp16")];
             tensor<string, []> query_33_pad_type_0 = const()[name = tensor<string, []>("query_33_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> query_33_strides_0 = const()[name = tensor<string, []>("query_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> query_33_pad_0 = const()[name = tensor<string, []>("query_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> query_33_dilations_0 = const()[name = tensor<string, []>("query_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> query_33_groups_0 = const()[name = tensor<string, []>("query_33_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_8_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(429262784)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_33_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = query_33_dilations_0, groups = query_33_groups_0, pad = query_33_pad_0, pad_type = query_33_pad_type_0, strides = query_33_strides_0, weight = layers_8_self_attn_q_proj_weight_to_fp16, x = obj_35_cast_fp16)[name = tensor<string, []>("query_33_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> query_33_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_33_dilations_0, groups = query_33_groups_0, pad = query_33_pad_0, pad_type = query_33_pad_type_0, strides = query_33_strides_0, weight = layers_8_self_attn_q_proj_weight_to_fp16, x = obj_35_cast_fp16)[name = tensor<string, []>("query_33_cast_fp16")];
             tensor<string, []> key_17_pad_type_0 = const()[name = tensor<string, []>("key_17_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> key_17_strides_0 = const()[name = tensor<string, []>("key_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> key_17_pad_0 = const()[name = tensor<string, []>("key_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
@@ -1552,173 +1554,173 @@ program(1.0)
             tensor<int32, [2]> value_17_dilations_0 = const()[name = tensor<string, []>("value_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> value_17_groups_0 = const()[name = tensor<string, []>("value_17_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_8_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(433457216)))];
-            tensor<fp16, [1, 1024, 1, 188]> value_17_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = value_17_dilations_0, groups = value_17_groups_0, pad = value_17_pad_0, pad_type = value_17_pad_type_0, strides = value_17_strides_0, weight = layers_8_self_attn_v_proj_weight_to_fp16, x = obj_35_cast_fp16)[name = tensor<string, []>("value_17_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_2496_to_fp16 = const()[name = tensor<string, []>("op_2496_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(435554432)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_35_cast_fp16 = add(x = query_33_cast_fp16, y = var_2496_to_fp16)[name = tensor<string, []>("query_35_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_2499_to_fp16 = const()[name = tensor<string, []>("op_2499_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(435556544)))];
-            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_17_cast_fp16 = add(x = query_33_cast_fp16, y = var_2499_to_fp16)[name = tensor<string, []>("q_with_bias_v_17_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> value_17_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_17_dilations_0, groups = value_17_groups_0, pad = value_17_pad_0, pad_type = value_17_pad_type_0, strides = value_17_strides_0, weight = layers_8_self_attn_v_proj_weight_to_fp16, x = obj_35_cast_fp16)[name = tensor<string, []>("value_17_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_2494_to_fp16 = const()[name = tensor<string, []>("op_2494_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(435554432)))];
+            tensor<fp16, [1, 1024, 1, 188]> query_35_cast_fp16 = add(x = query_33_cast_fp16, y = var_2494_to_fp16)[name = tensor<string, []>("query_35_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_2497_to_fp16 = const()[name = tensor<string, []>("op_2497_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(435556544)))];
+            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_17_cast_fp16 = add(x = query_33_cast_fp16, y = var_2497_to_fp16)[name = tensor<string, []>("q_with_bias_v_17_cast_fp16")];
             tensor<string, []> p_17_pad_type_0 = const()[name = tensor<string, []>("p_17_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> p_17_strides_0 = const()[name = tensor<string, []>("p_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> p_17_pad_0 = const()[name = tensor<string, []>("p_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> p_17_dilations_0 = const()[name = tensor<string, []>("p_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> p_17_groups_0 = const()[name = tensor<string, []>("p_17_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_8_self_attn_linear_pos_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_linear_pos_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(435558656)))];
-            tensor<fp16, [1, 1024, 1, 375]> p_17_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = p_17_dilations_0, groups = p_17_groups_0, pad = p_17_pad_0, pad_type = p_17_pad_type_0, strides = p_17_strides_0, weight = layers_8_self_attn_linear_pos_weight_to_fp16, x = pos_enc_to_fp16)[name = tensor<string, []>("p_17_cast_fp16")];
-            tensor<int32, [4]> var_2510 = const()[name = tensor<string, []>("op_2510"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_2511_cast_fp16 = reshape(shape = var_2510, x = q_with_bias_v_17_cast_fp16)[name = tensor<string, []>("op_2511_cast_fp16")];
-            tensor<int32, [4]> var_2512 = const()[name = tensor<string, []>("op_2512"), val = tensor<int32, [4]>([1, 8, 128, -1])];
-            tensor<fp16, [1, 8, 128, 375]> var_2513_cast_fp16 = reshape(shape = var_2512, x = p_17_cast_fp16)[name = tensor<string, []>("op_2513_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 375]> p_17_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_17_dilations_0, groups = p_17_groups_0, pad = p_17_pad_0, pad_type = p_17_pad_type_0, strides = p_17_strides_0, weight = layers_8_self_attn_linear_pos_weight_to_fp16, x = obj_3_cast_fp16)[name = tensor<string, []>("p_17_cast_fp16")];
+            tensor<int32, [4]> var_2508 = const()[name = tensor<string, []>("op_2508"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_2509_cast_fp16 = reshape(shape = var_2508, x = q_with_bias_v_17_cast_fp16)[name = tensor<string, []>("op_2509_cast_fp16")];
+            tensor<int32, [4]> var_2510 = const()[name = tensor<string, []>("op_2510"), val = tensor<int32, [4]>([1, 8, 128, -1])];
+            tensor<fp16, [1, 8, 128, 375]> var_2511_cast_fp16 = reshape(shape = var_2510, x = p_17_cast_fp16)[name = tensor<string, []>("op_2511_cast_fp16")];
             tensor<bool, []> matrix_bd_65_transpose_x_0 = const()[name = tensor<string, []>("matrix_bd_65_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> matrix_bd_65_transpose_y_0 = const()[name = tensor<string, []>("matrix_bd_65_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_65_cast_fp16 = matmul(transpose_x = matrix_bd_65_transpose_x_0, transpose_y = matrix_bd_65_transpose_y_0, x = var_2511_cast_fp16, y = var_2513_cast_fp16)[name = tensor<string, []>("matrix_bd_65_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_65_cast_fp16 = matmul(transpose_x = matrix_bd_65_transpose_x_0, transpose_y = matrix_bd_65_transpose_y_0, x = var_2509_cast_fp16, y = var_2511_cast_fp16)[name = tensor<string, []>("matrix_bd_65_cast_fp16")];
             tensor<int32, [8]> matrix_bd_67_pad_0 = const()[name = tensor<string, []>("matrix_bd_67_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
             tensor<string, []> matrix_bd_67_mode_0 = const()[name = tensor<string, []>("matrix_bd_67_mode_0"), val = tensor<string, []>("constant")];
             tensor<fp16, []> const_98_to_fp16 = const()[name = tensor<string, []>("const_98_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
             tensor<fp16, [1, 8, 188, 376]> matrix_bd_67_cast_fp16 = pad(constant_val = const_98_to_fp16, mode = matrix_bd_67_mode_0, pad = matrix_bd_67_pad_0, x = matrix_bd_65_cast_fp16)[name = tensor<string, []>("matrix_bd_67_cast_fp16")];
-            tensor<int32, [4]> var_2522 = const()[name = tensor<string, []>("op_2522"), val = tensor<int32, [4]>([1, 8, -1, 188])];
-            tensor<fp16, [1, 8, 376, 188]> matrix_bd_69_cast_fp16 = reshape(shape = var_2522, x = matrix_bd_67_cast_fp16)[name = tensor<string, []>("matrix_bd_69_cast_fp16")];
-            tensor<int32, [4]> var_2526_begin_0 = const()[name = tensor<string, []>("op_2526_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
-            tensor<int32, [4]> var_2526_end_0 = const()[name = tensor<string, []>("op_2526_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
-            tensor<bool, [4]> var_2526_end_mask_0 = const()[name = tensor<string, []>("op_2526_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
-            tensor<fp16, [1, 8, 375, 188]> var_2526_cast_fp16 = slice_by_index(begin = var_2526_begin_0, end = var_2526_end_0, end_mask = var_2526_end_mask_0, x = matrix_bd_69_cast_fp16)[name = tensor<string, []>("op_2526_cast_fp16")];
-            tensor<int32, [4]> var_2527 = const()[name = tensor<string, []>("op_2527"), val = tensor<int32, [4]>([1, 8, 188, 375])];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_71_cast_fp16 = reshape(shape = var_2527, x = var_2526_cast_fp16)[name = tensor<string, []>("matrix_bd_71_cast_fp16")];
-            tensor<int32, [4]> var_2532_begin_0 = const()[name = tensor<string, []>("op_2532_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [4]> var_2532_end_0 = const()[name = tensor<string, []>("op_2532_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
-            tensor<bool, [4]> var_2532_end_mask_0 = const()[name = tensor<string, []>("op_2532_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
-            tensor<fp16, [1, 8, 188, 188]> var_2532_cast_fp16 = slice_by_index(begin = var_2532_begin_0, end = var_2532_end_0, end_mask = var_2532_end_mask_0, x = matrix_bd_71_cast_fp16)[name = tensor<string, []>("op_2532_cast_fp16")];
-            tensor<fp16, []> var_2533_to_fp16 = const()[name = tensor<string, []>("op_2533_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 188, 188]> qk_mask_17_cast_fp16 = mul(x = var_2532_cast_fp16, y = var_2533_to_fp16)[name = tensor<string, []>("qk_mask_17_cast_fp16")];
-            tensor<int32, [4]> var_2537 = const()[name = tensor<string, []>("op_2537"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> mh_q_17_cast_fp16 = reshape(shape = var_2537, x = query_35_cast_fp16)[name = tensor<string, []>("mh_q_17_cast_fp16")];
-            tensor<fp16, []> var_2539_to_fp16 = const()[name = tensor<string, []>("op_2539_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 128, 188]> var_2540_cast_fp16 = mul(x = mh_q_17_cast_fp16, y = var_2539_to_fp16)[name = tensor<string, []>("op_2540_cast_fp16")];
-            tensor<int32, [4]> var_2543 = const()[name = tensor<string, []>("op_2543"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_2544_cast_fp16 = reshape(shape = var_2543, x = key_17_cast_fp16)[name = tensor<string, []>("op_2544_cast_fp16")];
+            tensor<int32, [4]> var_2520 = const()[name = tensor<string, []>("op_2520"), val = tensor<int32, [4]>([1, 8, -1, 188])];
+            tensor<fp16, [1, 8, 376, 188]> matrix_bd_69_cast_fp16 = reshape(shape = var_2520, x = matrix_bd_67_cast_fp16)[name = tensor<string, []>("matrix_bd_69_cast_fp16")];
+            tensor<int32, [4]> var_2524_begin_0 = const()[name = tensor<string, []>("op_2524_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2524_end_0 = const()[name = tensor<string, []>("op_2524_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
+            tensor<bool, [4]> var_2524_end_mask_0 = const()[name = tensor<string, []>("op_2524_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 375, 188]> var_2524_cast_fp16 = slice_by_index(begin = var_2524_begin_0, end = var_2524_end_0, end_mask = var_2524_end_mask_0, x = matrix_bd_69_cast_fp16)[name = tensor<string, []>("op_2524_cast_fp16")];
+            tensor<int32, [4]> var_2525 = const()[name = tensor<string, []>("op_2525"), val = tensor<int32, [4]>([1, 8, 188, 375])];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_71_cast_fp16 = reshape(shape = var_2525, x = var_2524_cast_fp16)[name = tensor<string, []>("matrix_bd_71_cast_fp16")];
+            tensor<int32, [4]> var_2530_begin_0 = const()[name = tensor<string, []>("op_2530_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_2530_end_0 = const()[name = tensor<string, []>("op_2530_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
+            tensor<bool, [4]> var_2530_end_mask_0 = const()[name = tensor<string, []>("op_2530_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 188, 188]> var_2530_cast_fp16 = slice_by_index(begin = var_2530_begin_0, end = var_2530_end_0, end_mask = var_2530_end_mask_0, x = matrix_bd_71_cast_fp16)[name = tensor<string, []>("op_2530_cast_fp16")];
+            tensor<fp16, []> var_2531_to_fp16 = const()[name = tensor<string, []>("op_2531_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 188, 188]> qk_mask_17_cast_fp16 = mul(x = var_2530_cast_fp16, y = var_2531_to_fp16)[name = tensor<string, []>("qk_mask_17_cast_fp16")];
+            tensor<int32, [4]> var_2535 = const()[name = tensor<string, []>("op_2535"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> mh_q_17_cast_fp16 = reshape(shape = var_2535, x = query_35_cast_fp16)[name = tensor<string, []>("mh_q_17_cast_fp16")];
+            tensor<fp16, []> var_2537_to_fp16 = const()[name = tensor<string, []>("op_2537_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 128, 188]> var_2538_cast_fp16 = mul(x = mh_q_17_cast_fp16, y = var_2537_to_fp16)[name = tensor<string, []>("op_2538_cast_fp16")];
+            tensor<int32, [4]> var_2541 = const()[name = tensor<string, []>("op_2541"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_2542_cast_fp16 = reshape(shape = var_2541, x = key_17_cast_fp16)[name = tensor<string, []>("op_2542_cast_fp16")];
             tensor<bool, []> mh_w_33_transpose_x_0 = const()[name = tensor<string, []>("mh_w_33_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> mh_w_33_transpose_y_0 = const()[name = tensor<string, []>("mh_w_33_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 188]> mh_w_33_cast_fp16 = matmul(transpose_x = mh_w_33_transpose_x_0, transpose_y = mh_w_33_transpose_y_0, x = var_2540_cast_fp16, y = var_2544_cast_fp16)[name = tensor<string, []>("mh_w_33_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> mh_w_33_cast_fp16 = matmul(transpose_x = mh_w_33_transpose_x_0, transpose_y = mh_w_33_transpose_y_0, x = var_2538_cast_fp16, y = var_2542_cast_fp16)[name = tensor<string, []>("mh_w_33_cast_fp16")];
             tensor<fp16, [1, 8, 188, 188]> mh_w_35_cast_fp16 = add(x = mh_w_33_cast_fp16, y = qk_mask_17_cast_fp16)[name = tensor<string, []>("mh_w_35_cast_fp16")];
-            tensor<fp16, [1, 8, 188, 188]> var_2548_cast_fp16 = softmax(axis = var_2389, x = mh_w_35_cast_fp16)[name = tensor<string, []>("op_2548_cast_fp16")];
-            tensor<int32, [4]> var_2549 = const()[name = tensor<string, []>("op_2549"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_2550_cast_fp16 = reshape(shape = var_2549, x = value_17_cast_fp16)[name = tensor<string, []>("op_2550_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> var_2546_cast_fp16 = softmax(axis = var_2387, x = mh_w_35_cast_fp16)[name = tensor<string, []>("op_2546_cast_fp16")];
+            tensor<int32, [4]> var_2547 = const()[name = tensor<string, []>("op_2547"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_2548_cast_fp16 = reshape(shape = var_2547, x = value_17_cast_fp16)[name = tensor<string, []>("op_2548_cast_fp16")];
             tensor<bool, []> attn_17_transpose_x_0 = const()[name = tensor<string, []>("attn_17_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> attn_17_transpose_y_0 = const()[name = tensor<string, []>("attn_17_transpose_y_0"), val = tensor<bool, []>(true)];
-            tensor<fp16, [1, 8, 128, 188]> attn_17_cast_fp16 = matmul(transpose_x = attn_17_transpose_x_0, transpose_y = attn_17_transpose_y_0, x = var_2550_cast_fp16, y = var_2548_cast_fp16)[name = tensor<string, []>("attn_17_cast_fp16")];
-            tensor<int32, [4]> var_2553 = const()[name = tensor<string, []>("op_2553"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
-            tensor<fp16, [1, 1024, 1, 188]> input_229_cast_fp16 = reshape(shape = var_2553, x = attn_17_cast_fp16)[name = tensor<string, []>("input_229_cast_fp16")];
+            tensor<fp16, [1, 8, 128, 188]> attn_17_cast_fp16 = matmul(transpose_x = attn_17_transpose_x_0, transpose_y = attn_17_transpose_y_0, x = var_2548_cast_fp16, y = var_2546_cast_fp16)[name = tensor<string, []>("attn_17_cast_fp16")];
+            tensor<int32, [4]> var_2551 = const()[name = tensor<string, []>("op_2551"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
+            tensor<fp16, [1, 1024, 1, 188]> input_231_cast_fp16 = reshape(shape = var_2551, x = attn_17_cast_fp16)[name = tensor<string, []>("input_231_cast_fp16")];
             tensor<string, []> obj_37_pad_type_0 = const()[name = tensor<string, []>("obj_37_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> obj_37_strides_0 = const()[name = tensor<string, []>("obj_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> obj_37_pad_0 = const()[name = tensor<string, []>("obj_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> obj_37_dilations_0 = const()[name = tensor<string, []>("obj_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> obj_37_groups_0 = const()[name = tensor<string, []>("obj_37_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_8_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(437655872)))];
-            tensor<fp16, [1, 1024, 1, 188]> obj_37_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = obj_37_dilations_0, groups = obj_37_groups_0, pad = obj_37_pad_0, pad_type = obj_37_pad_type_0, strides = obj_37_strides_0, weight = layers_8_self_attn_o_proj_weight_to_fp16, x = input_229_cast_fp16)[name = tensor<string, []>("obj_37_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_37_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_37_dilations_0, groups = obj_37_groups_0, pad = obj_37_pad_0, pad_type = obj_37_pad_type_0, strides = obj_37_strides_0, weight = layers_8_self_attn_o_proj_weight_to_fp16, x = input_231_cast_fp16)[name = tensor<string, []>("obj_37_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_85_cast_fp16 = add(x = inputs_83_cast_fp16, y = obj_37_cast_fp16)[name = tensor<string, []>("inputs_85_cast_fp16")];
             tensor<int32, [1]> out_85_axes_0 = const()[name = tensor<string, []>("out_85_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_2571_to_fp16 = const()[name = tensor<string, []>("op_2571_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_85_cast_fp16 = layer_norm(axes = out_85_axes_0, epsilon = var_2571_to_fp16, x = inputs_85_cast_fp16)[name = tensor<string, []>("out_85_cast_fp16")];
-            tensor<fp16, [1024]> input_231_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_231_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(439753088)))];
-            tensor<fp16, [1024]> input_231_beta_0_to_fp16 = const()[name = tensor<string, []>("input_231_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(439755200)))];
-            tensor<fp16, []> input_231_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_231_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_231_cast_fp16 = batch_norm(beta = input_231_beta_0_to_fp16, epsilon = input_231_epsilon_0_to_fp16, gamma = input_231_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_85_cast_fp16)[name = tensor<string, []>("input_231_cast_fp16")];
-            tensor<string, []> input_233_pad_type_0 = const()[name = tensor<string, []>("input_233_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_233_strides_0 = const()[name = tensor<string, []>("input_233_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_233_pad_0 = const()[name = tensor<string, []>("input_233_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_233_dilations_0 = const()[name = tensor<string, []>("input_233_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_233_groups_0 = const()[name = tensor<string, []>("input_233_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_2569_to_fp16 = const()[name = tensor<string, []>("op_2569_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_85_cast_fp16 = layer_norm(axes = out_85_axes_0, epsilon = var_2569_to_fp16, x = inputs_85_cast_fp16)[name = tensor<string, []>("out_85_cast_fp16")];
+            tensor<fp16, [1024]> input_233_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_233_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(439753088)))];
+            tensor<fp16, [1024]> input_233_beta_0_to_fp16 = const()[name = tensor<string, []>("input_233_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(439755200)))];
+            tensor<fp16, []> input_233_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_233_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_233_cast_fp16 = batch_norm(beta = input_233_beta_0_to_fp16, epsilon = input_233_epsilon_0_to_fp16, gamma = input_233_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_85_cast_fp16)[name = tensor<string, []>("input_233_cast_fp16")];
+            tensor<string, []> input_235_pad_type_0 = const()[name = tensor<string, []>("input_235_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_235_strides_0 = const()[name = tensor<string, []>("input_235_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_235_pad_0 = const()[name = tensor<string, []>("input_235_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_235_dilations_0 = const()[name = tensor<string, []>("input_235_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_235_groups_0 = const()[name = tensor<string, []>("input_235_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [2048, 1024, 1, 1]> layers_8_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(439757312)))];
-            tensor<fp16, [1, 2048, 1, 188]> input_233_cast_fp16 = conv(dilations = input_233_dilations_0, groups = input_233_groups_0, pad = input_233_pad_0, pad_type = input_233_pad_type_0, strides = input_233_strides_0, weight = layers_8_conv_pointwise_conv1_weight_to_fp16, x = input_231_cast_fp16)[name = tensor<string, []>("input_233_cast_fp16")];
-            tensor<int32, []> input_235_split_num_splits_0 = const()[name = tensor<string, []>("input_235_split_num_splits_0"), val = tensor<int32, []>(2)];
-            tensor<int32, []> input_235_split_axis_0 = const()[name = tensor<string, []>("input_235_split_axis_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1, 1024, 1, 188]> input_235_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_235_split_cast_fp16_1 = split(axis = input_235_split_axis_0, num_splits = input_235_split_num_splits_0, x = input_233_cast_fp16)[name = tensor<string, []>("input_235_split_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_235_split_1_sigmoid_cast_fp16 = sigmoid(x = input_235_split_cast_fp16_1)[name = tensor<string, []>("input_235_split_1_sigmoid_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_235_cast_fp16 = mul(x = input_235_split_cast_fp16_0, y = input_235_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_235_cast_fp16")];
-            tensor<string, []> input_237_pad_type_0 = const()[name = tensor<string, []>("input_237_pad_type_0"), val = tensor<string, []>("custom")];
-            tensor<int32, [4]> input_237_pad_0 = const()[name = tensor<string, []>("input_237_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
-            tensor<int32, []> input_237_groups_0 = const()[name = tensor<string, []>("input_237_groups_0"), val = tensor<int32, []>(1024)];
-            tensor<int32, [2]> input_237_strides_0 = const()[name = tensor<string, []>("input_237_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> input_237_dilations_0 = const()[name = tensor<string, []>("input_237_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [1, 2048, 1, 188]> input_235_cast_fp16 = conv(dilations = input_235_dilations_0, groups = input_235_groups_0, pad = input_235_pad_0, pad_type = input_235_pad_type_0, strides = input_235_strides_0, weight = layers_8_conv_pointwise_conv1_weight_to_fp16, x = input_233_cast_fp16)[name = tensor<string, []>("input_235_cast_fp16")];
+            tensor<int32, []> input_237_split_num_splits_0 = const()[name = tensor<string, []>("input_237_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> input_237_split_axis_0 = const()[name = tensor<string, []>("input_237_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 1, 188]> input_237_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_237_split_cast_fp16_1 = split(axis = input_237_split_axis_0, num_splits = input_237_split_num_splits_0, x = input_235_cast_fp16)[name = tensor<string, []>("input_237_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_237_split_1_sigmoid_cast_fp16 = sigmoid(x = input_237_split_cast_fp16_1)[name = tensor<string, []>("input_237_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_237_cast_fp16 = mul(x = input_237_split_cast_fp16_0, y = input_237_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_237_cast_fp16")];
+            tensor<string, []> input_239_pad_type_0 = const()[name = tensor<string, []>("input_239_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_239_pad_0 = const()[name = tensor<string, []>("input_239_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
+            tensor<int32, []> input_239_groups_0 = const()[name = tensor<string, []>("input_239_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [2]> input_239_strides_0 = const()[name = tensor<string, []>("input_239_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> input_239_dilations_0 = const()[name = tensor<string, []>("input_239_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<fp16, [1024, 1, 1, 9]> const_284_to_fp16 = const()[name = tensor<string, []>("const_284_to_fp16"), val = tensor<fp16, [1024, 1, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(443951680)))];
             tensor<fp16, [1024]> const_285_to_fp16 = const()[name = tensor<string, []>("const_285_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(443970176)))];
-            tensor<fp16, [1, 1024, 1, 188]> input_239_cast_fp16 = conv(bias = const_285_to_fp16, dilations = input_237_dilations_0, groups = input_237_groups_0, pad = input_237_pad_0, pad_type = input_237_pad_type_0, strides = input_237_strides_0, weight = const_284_to_fp16, x = input_235_cast_fp16)[name = tensor<string, []>("input_239_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_241_cast_fp16 = silu(x = input_239_cast_fp16)[name = tensor<string, []>("input_241_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_241_cast_fp16 = conv(bias = const_285_to_fp16, dilations = input_239_dilations_0, groups = input_239_groups_0, pad = input_239_pad_0, pad_type = input_239_pad_type_0, strides = input_239_strides_0, weight = const_284_to_fp16, x = input_237_cast_fp16)[name = tensor<string, []>("input_241_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_243_cast_fp16 = silu(x = input_241_cast_fp16)[name = tensor<string, []>("input_243_cast_fp16")];
             tensor<string, []> x_53_pad_type_0 = const()[name = tensor<string, []>("x_53_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_53_strides_0 = const()[name = tensor<string, []>("x_53_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_53_pad_0 = const()[name = tensor<string, []>("x_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_53_dilations_0 = const()[name = tensor<string, []>("x_53_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_53_groups_0 = const()[name = tensor<string, []>("x_53_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_8_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(443972288)))];
-            tensor<fp16, [1, 1024, 1, 188]> x_53_cast_fp16 = conv(dilations = x_53_dilations_0, groups = x_53_groups_0, pad = x_53_pad_0, pad_type = x_53_pad_type_0, strides = x_53_strides_0, weight = layers_8_conv_pointwise_conv2_weight_to_fp16, x = input_241_cast_fp16)[name = tensor<string, []>("x_53_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> x_53_cast_fp16 = conv(dilations = x_53_dilations_0, groups = x_53_groups_0, pad = x_53_pad_0, pad_type = x_53_pad_type_0, strides = x_53_strides_0, weight = layers_8_conv_pointwise_conv2_weight_to_fp16, x = input_243_cast_fp16)[name = tensor<string, []>("x_53_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_87_cast_fp16 = add(x = inputs_85_cast_fp16, y = x_53_cast_fp16)[name = tensor<string, []>("inputs_87_cast_fp16")];
             tensor<int32, [1]> out_87_axes_0 = const()[name = tensor<string, []>("out_87_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_2619_to_fp16 = const()[name = tensor<string, []>("op_2619_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_87_cast_fp16 = layer_norm(axes = out_87_axes_0, epsilon = var_2619_to_fp16, x = inputs_87_cast_fp16)[name = tensor<string, []>("out_87_cast_fp16")];
-            tensor<fp16, [1024]> input_243_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_243_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(446069504)))];
-            tensor<fp16, [1024]> input_243_beta_0_to_fp16 = const()[name = tensor<string, []>("input_243_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(446071616)))];
-            tensor<fp16, []> input_243_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_243_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_243_cast_fp16 = batch_norm(beta = input_243_beta_0_to_fp16, epsilon = input_243_epsilon_0_to_fp16, gamma = input_243_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_87_cast_fp16)[name = tensor<string, []>("input_243_cast_fp16")];
-            tensor<string, []> input_245_pad_type_0 = const()[name = tensor<string, []>("input_245_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_245_strides_0 = const()[name = tensor<string, []>("input_245_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_245_pad_0 = const()[name = tensor<string, []>("input_245_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_245_dilations_0 = const()[name = tensor<string, []>("input_245_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_245_groups_0 = const()[name = tensor<string, []>("input_245_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_2617_to_fp16 = const()[name = tensor<string, []>("op_2617_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_87_cast_fp16 = layer_norm(axes = out_87_axes_0, epsilon = var_2617_to_fp16, x = inputs_87_cast_fp16)[name = tensor<string, []>("out_87_cast_fp16")];
+            tensor<fp16, [1024]> input_245_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_245_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(446069504)))];
+            tensor<fp16, [1024]> input_245_beta_0_to_fp16 = const()[name = tensor<string, []>("input_245_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(446071616)))];
+            tensor<fp16, []> input_245_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_245_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_245_cast_fp16 = batch_norm(beta = input_245_beta_0_to_fp16, epsilon = input_245_epsilon_0_to_fp16, gamma = input_245_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_87_cast_fp16)[name = tensor<string, []>("input_245_cast_fp16")];
+            tensor<string, []> input_247_pad_type_0 = const()[name = tensor<string, []>("input_247_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_247_strides_0 = const()[name = tensor<string, []>("input_247_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_247_pad_0 = const()[name = tensor<string, []>("input_247_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_247_dilations_0 = const()[name = tensor<string, []>("input_247_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_247_groups_0 = const()[name = tensor<string, []>("input_247_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_8_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_feed_forward2_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(446073728)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_245_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_245_dilations_0, groups = input_245_groups_0, pad = input_245_pad_0, pad_type = input_245_pad_type_0, strides = input_245_strides_0, weight = layers_8_feed_forward2_fc1_weight_to_fp16, x = input_243_cast_fp16)[name = tensor<string, []>("input_245_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_247_cast_fp16 = silu(x = input_245_cast_fp16)[name = tensor<string, []>("input_247_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_247_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_247_dilations_0, groups = input_247_groups_0, pad = input_247_pad_0, pad_type = input_247_pad_type_0, strides = input_247_strides_0, weight = layers_8_feed_forward2_fc1_weight_to_fp16, x = input_245_cast_fp16)[name = tensor<string, []>("input_247_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_249_cast_fp16 = silu(x = input_247_cast_fp16)[name = tensor<string, []>("input_249_cast_fp16")];
             tensor<string, []> x_55_pad_type_0 = const()[name = tensor<string, []>("x_55_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_55_strides_0 = const()[name = tensor<string, []>("x_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_55_pad_0 = const()[name = tensor<string, []>("x_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_55_dilations_0 = const()[name = tensor<string, []>("x_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_55_groups_0 = const()[name = tensor<string, []>("x_55_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_2647_weight_0_to_fp16 = const()[name = tensor<string, []>("op_2647_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(454462400)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_2647_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_55_dilations_0, groups = x_55_groups_0, pad = x_55_pad_0, pad_type = x_55_pad_type_0, strides = x_55_strides_0, weight = var_2647_weight_0_to_fp16, x = input_247_cast_fp16)[name = tensor<string, []>("op_2647_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_89_cast_fp16 = add(x = inputs_87_cast_fp16, y = var_2647_cast_fp16)[name = tensor<string, []>("inputs_89_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_2645_weight_0_to_fp16 = const()[name = tensor<string, []>("op_2645_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(454462400)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_2645_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_55_dilations_0, groups = x_55_groups_0, pad = x_55_pad_0, pad_type = x_55_pad_type_0, strides = x_55_strides_0, weight = var_2645_weight_0_to_fp16, x = input_249_cast_fp16)[name = tensor<string, []>("op_2645_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_89_cast_fp16 = add(x = inputs_87_cast_fp16, y = var_2645_cast_fp16)[name = tensor<string, []>("inputs_89_cast_fp16")];
             tensor<int32, [1]> out_89_axes_0 = const()[name = tensor<string, []>("out_89_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_2657_to_fp16 = const()[name = tensor<string, []>("op_2657_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_89_cast_fp16 = layer_norm(axes = out_89_axes_0, epsilon = var_2657_to_fp16, x = inputs_89_cast_fp16)[name = tensor<string, []>("out_89_cast_fp16")];
+            tensor<fp16, []> var_2655_to_fp16 = const()[name = tensor<string, []>("op_2655_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_89_cast_fp16 = layer_norm(axes = out_89_axes_0, epsilon = var_2655_to_fp16, x = inputs_89_cast_fp16)[name = tensor<string, []>("out_89_cast_fp16")];
             tensor<fp16, [1024]> inputs_91_gamma_0_to_fp16 = const()[name = tensor<string, []>("inputs_91_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(462851072)))];
             tensor<fp16, [1024]> inputs_91_beta_0_to_fp16 = const()[name = tensor<string, []>("inputs_91_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(462853184)))];
             tensor<fp16, []> inputs_91_epsilon_0_to_fp16 = const()[name = tensor<string, []>("inputs_91_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_91_cast_fp16 = batch_norm(beta = inputs_91_beta_0_to_fp16, epsilon = inputs_91_epsilon_0_to_fp16, gamma = inputs_91_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_89_cast_fp16)[name = tensor<string, []>("inputs_91_cast_fp16")];
-            tensor<int32, []> var_2671 = const()[name = tensor<string, []>("op_2671"), val = tensor<int32, []>(3)];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_91_cast_fp16 = batch_norm(beta = inputs_91_beta_0_to_fp16, epsilon = inputs_91_epsilon_0_to_fp16, gamma = inputs_91_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_89_cast_fp16)[name = tensor<string, []>("inputs_91_cast_fp16")];
+            tensor<int32, []> var_2669 = const()[name = tensor<string, []>("op_2669"), val = tensor<int32, []>(3)];
             tensor<int32, [1]> out_91_axes_0 = const()[name = tensor<string, []>("out_91_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_2702_to_fp16 = const()[name = tensor<string, []>("op_2702_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_91_cast_fp16 = layer_norm(axes = out_91_axes_0, epsilon = var_2702_to_fp16, x = inputs_91_cast_fp16)[name = tensor<string, []>("out_91_cast_fp16")];
-            tensor<fp16, [1024]> input_249_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_249_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(462855296)))];
-            tensor<fp16, [1024]> input_249_beta_0_to_fp16 = const()[name = tensor<string, []>("input_249_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(462857408)))];
-            tensor<fp16, []> input_249_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_249_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_249_cast_fp16 = batch_norm(beta = input_249_beta_0_to_fp16, epsilon = input_249_epsilon_0_to_fp16, gamma = input_249_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_91_cast_fp16)[name = tensor<string, []>("input_249_cast_fp16")];
-            tensor<string, []> input_251_pad_type_0 = const()[name = tensor<string, []>("input_251_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_251_strides_0 = const()[name = tensor<string, []>("input_251_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_251_pad_0 = const()[name = tensor<string, []>("input_251_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_251_dilations_0 = const()[name = tensor<string, []>("input_251_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_251_groups_0 = const()[name = tensor<string, []>("input_251_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_2700_to_fp16 = const()[name = tensor<string, []>("op_2700_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_91_cast_fp16 = layer_norm(axes = out_91_axes_0, epsilon = var_2700_to_fp16, x = inputs_91_cast_fp16)[name = tensor<string, []>("out_91_cast_fp16")];
+            tensor<fp16, [1024]> input_251_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_251_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(462855296)))];
+            tensor<fp16, [1024]> input_251_beta_0_to_fp16 = const()[name = tensor<string, []>("input_251_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(462857408)))];
+            tensor<fp16, []> input_251_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_251_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_251_cast_fp16 = batch_norm(beta = input_251_beta_0_to_fp16, epsilon = input_251_epsilon_0_to_fp16, gamma = input_251_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_91_cast_fp16)[name = tensor<string, []>("input_251_cast_fp16")];
+            tensor<string, []> input_253_pad_type_0 = const()[name = tensor<string, []>("input_253_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_253_strides_0 = const()[name = tensor<string, []>("input_253_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_253_pad_0 = const()[name = tensor<string, []>("input_253_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_253_dilations_0 = const()[name = tensor<string, []>("input_253_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_253_groups_0 = const()[name = tensor<string, []>("input_253_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_9_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_feed_forward1_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(462859520)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_251_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_251_dilations_0, groups = input_251_groups_0, pad = input_251_pad_0, pad_type = input_251_pad_type_0, strides = input_251_strides_0, weight = layers_9_feed_forward1_fc1_weight_to_fp16, x = input_249_cast_fp16)[name = tensor<string, []>("input_251_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_253_cast_fp16 = silu(x = input_251_cast_fp16)[name = tensor<string, []>("input_253_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_253_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_253_dilations_0, groups = input_253_groups_0, pad = input_253_pad_0, pad_type = input_253_pad_type_0, strides = input_253_strides_0, weight = layers_9_feed_forward1_fc1_weight_to_fp16, x = input_251_cast_fp16)[name = tensor<string, []>("input_253_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_255_cast_fp16 = silu(x = input_253_cast_fp16)[name = tensor<string, []>("input_255_cast_fp16")];
             tensor<string, []> x_57_pad_type_0 = const()[name = tensor<string, []>("x_57_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_57_strides_0 = const()[name = tensor<string, []>("x_57_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_57_pad_0 = const()[name = tensor<string, []>("x_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_57_dilations_0 = const()[name = tensor<string, []>("x_57_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_57_groups_0 = const()[name = tensor<string, []>("x_57_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_2730_weight_0_to_fp16 = const()[name = tensor<string, []>("op_2730_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(471248192)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_2730_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_57_dilations_0, groups = x_57_groups_0, pad = x_57_pad_0, pad_type = x_57_pad_type_0, strides = x_57_strides_0, weight = var_2730_weight_0_to_fp16, x = input_253_cast_fp16)[name = tensor<string, []>("op_2730_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_93_cast_fp16 = add(x = inputs_91_cast_fp16, y = var_2730_cast_fp16)[name = tensor<string, []>("inputs_93_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_2728_weight_0_to_fp16 = const()[name = tensor<string, []>("op_2728_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(471248192)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_2728_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_57_dilations_0, groups = x_57_groups_0, pad = x_57_pad_0, pad_type = x_57_pad_type_0, strides = x_57_strides_0, weight = var_2728_weight_0_to_fp16, x = input_255_cast_fp16)[name = tensor<string, []>("op_2728_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_93_cast_fp16 = add(x = inputs_91_cast_fp16, y = var_2728_cast_fp16)[name = tensor<string, []>("inputs_93_cast_fp16")];
             tensor<int32, [1]> out_93_axes_0 = const()[name = tensor<string, []>("out_93_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_2740_to_fp16 = const()[name = tensor<string, []>("op_2740_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_93_cast_fp16 = layer_norm(axes = out_93_axes_0, epsilon = var_2740_to_fp16, x = inputs_93_cast_fp16)[name = tensor<string, []>("out_93_cast_fp16")];
+            tensor<fp16, []> var_2738_to_fp16 = const()[name = tensor<string, []>("op_2738_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_93_cast_fp16 = layer_norm(axes = out_93_axes_0, epsilon = var_2738_to_fp16, x = inputs_93_cast_fp16)[name = tensor<string, []>("out_93_cast_fp16")];
             tensor<fp16, [1024]> obj_39_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_39_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(479636864)))];
             tensor<fp16, [1024]> obj_39_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_39_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(479638976)))];
             tensor<fp16, []> obj_39_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_39_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> obj_39_cast_fp16 = batch_norm(beta = obj_39_beta_0_to_fp16, epsilon = obj_39_epsilon_0_to_fp16, gamma = obj_39_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_93_cast_fp16)[name = tensor<string, []>("obj_39_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_39_cast_fp16 = batch_norm(beta = obj_39_beta_0_to_fp16, epsilon = obj_39_epsilon_0_to_fp16, gamma = obj_39_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_93_cast_fp16)[name = tensor<string, []>("obj_39_cast_fp16")];
             tensor<string, []> query_37_pad_type_0 = const()[name = tensor<string, []>("query_37_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> query_37_strides_0 = const()[name = tensor<string, []>("query_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> query_37_pad_0 = const()[name = tensor<string, []>("query_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> query_37_dilations_0 = const()[name = tensor<string, []>("query_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> query_37_groups_0 = const()[name = tensor<string, []>("query_37_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_9_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(479641088)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_37_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = query_37_dilations_0, groups = query_37_groups_0, pad = query_37_pad_0, pad_type = query_37_pad_type_0, strides = query_37_strides_0, weight = layers_9_self_attn_q_proj_weight_to_fp16, x = obj_39_cast_fp16)[name = tensor<string, []>("query_37_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> query_37_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_37_dilations_0, groups = query_37_groups_0, pad = query_37_pad_0, pad_type = query_37_pad_type_0, strides = query_37_strides_0, weight = layers_9_self_attn_q_proj_weight_to_fp16, x = obj_39_cast_fp16)[name = tensor<string, []>("query_37_cast_fp16")];
             tensor<string, []> key_19_pad_type_0 = const()[name = tensor<string, []>("key_19_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> key_19_strides_0 = const()[name = tensor<string, []>("key_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> key_19_pad_0 = const()[name = tensor<string, []>("key_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
@@ -1732,173 +1734,173 @@ program(1.0)
             tensor<int32, [2]> value_19_dilations_0 = const()[name = tensor<string, []>("value_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> value_19_groups_0 = const()[name = tensor<string, []>("value_19_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_9_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(483835520)))];
-            tensor<fp16, [1, 1024, 1, 188]> value_19_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = value_19_dilations_0, groups = value_19_groups_0, pad = value_19_pad_0, pad_type = value_19_pad_type_0, strides = value_19_strides_0, weight = layers_9_self_attn_v_proj_weight_to_fp16, x = obj_39_cast_fp16)[name = tensor<string, []>("value_19_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_2778_to_fp16 = const()[name = tensor<string, []>("op_2778_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(485932736)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_39_cast_fp16 = add(x = query_37_cast_fp16, y = var_2778_to_fp16)[name = tensor<string, []>("query_39_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_2781_to_fp16 = const()[name = tensor<string, []>("op_2781_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(485934848)))];
-            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_19_cast_fp16 = add(x = query_37_cast_fp16, y = var_2781_to_fp16)[name = tensor<string, []>("q_with_bias_v_19_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> value_19_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_19_dilations_0, groups = value_19_groups_0, pad = value_19_pad_0, pad_type = value_19_pad_type_0, strides = value_19_strides_0, weight = layers_9_self_attn_v_proj_weight_to_fp16, x = obj_39_cast_fp16)[name = tensor<string, []>("value_19_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_2776_to_fp16 = const()[name = tensor<string, []>("op_2776_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(485932736)))];
+            tensor<fp16, [1, 1024, 1, 188]> query_39_cast_fp16 = add(x = query_37_cast_fp16, y = var_2776_to_fp16)[name = tensor<string, []>("query_39_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_2779_to_fp16 = const()[name = tensor<string, []>("op_2779_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(485934848)))];
+            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_19_cast_fp16 = add(x = query_37_cast_fp16, y = var_2779_to_fp16)[name = tensor<string, []>("q_with_bias_v_19_cast_fp16")];
             tensor<string, []> p_19_pad_type_0 = const()[name = tensor<string, []>("p_19_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> p_19_strides_0 = const()[name = tensor<string, []>("p_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> p_19_pad_0 = const()[name = tensor<string, []>("p_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> p_19_dilations_0 = const()[name = tensor<string, []>("p_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> p_19_groups_0 = const()[name = tensor<string, []>("p_19_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_9_self_attn_linear_pos_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_linear_pos_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(485936960)))];
-            tensor<fp16, [1, 1024, 1, 375]> p_19_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = p_19_dilations_0, groups = p_19_groups_0, pad = p_19_pad_0, pad_type = p_19_pad_type_0, strides = p_19_strides_0, weight = layers_9_self_attn_linear_pos_weight_to_fp16, x = pos_enc_to_fp16)[name = tensor<string, []>("p_19_cast_fp16")];
-            tensor<int32, [4]> var_2792 = const()[name = tensor<string, []>("op_2792"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_2793_cast_fp16 = reshape(shape = var_2792, x = q_with_bias_v_19_cast_fp16)[name = tensor<string, []>("op_2793_cast_fp16")];
-            tensor<int32, [4]> var_2794 = const()[name = tensor<string, []>("op_2794"), val = tensor<int32, [4]>([1, 8, 128, -1])];
-            tensor<fp16, [1, 8, 128, 375]> var_2795_cast_fp16 = reshape(shape = var_2794, x = p_19_cast_fp16)[name = tensor<string, []>("op_2795_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 375]> p_19_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_19_dilations_0, groups = p_19_groups_0, pad = p_19_pad_0, pad_type = p_19_pad_type_0, strides = p_19_strides_0, weight = layers_9_self_attn_linear_pos_weight_to_fp16, x = obj_3_cast_fp16)[name = tensor<string, []>("p_19_cast_fp16")];
+            tensor<int32, [4]> var_2790 = const()[name = tensor<string, []>("op_2790"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_2791_cast_fp16 = reshape(shape = var_2790, x = q_with_bias_v_19_cast_fp16)[name = tensor<string, []>("op_2791_cast_fp16")];
+            tensor<int32, [4]> var_2792 = const()[name = tensor<string, []>("op_2792"), val = tensor<int32, [4]>([1, 8, 128, -1])];
+            tensor<fp16, [1, 8, 128, 375]> var_2793_cast_fp16 = reshape(shape = var_2792, x = p_19_cast_fp16)[name = tensor<string, []>("op_2793_cast_fp16")];
             tensor<bool, []> matrix_bd_73_transpose_x_0 = const()[name = tensor<string, []>("matrix_bd_73_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> matrix_bd_73_transpose_y_0 = const()[name = tensor<string, []>("matrix_bd_73_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_73_cast_fp16 = matmul(transpose_x = matrix_bd_73_transpose_x_0, transpose_y = matrix_bd_73_transpose_y_0, x = var_2793_cast_fp16, y = var_2795_cast_fp16)[name = tensor<string, []>("matrix_bd_73_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_73_cast_fp16 = matmul(transpose_x = matrix_bd_73_transpose_x_0, transpose_y = matrix_bd_73_transpose_y_0, x = var_2791_cast_fp16, y = var_2793_cast_fp16)[name = tensor<string, []>("matrix_bd_73_cast_fp16")];
             tensor<int32, [8]> matrix_bd_75_pad_0 = const()[name = tensor<string, []>("matrix_bd_75_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
             tensor<string, []> matrix_bd_75_mode_0 = const()[name = tensor<string, []>("matrix_bd_75_mode_0"), val = tensor<string, []>("constant")];
             tensor<fp16, []> const_109_to_fp16 = const()[name = tensor<string, []>("const_109_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
             tensor<fp16, [1, 8, 188, 376]> matrix_bd_75_cast_fp16 = pad(constant_val = const_109_to_fp16, mode = matrix_bd_75_mode_0, pad = matrix_bd_75_pad_0, x = matrix_bd_73_cast_fp16)[name = tensor<string, []>("matrix_bd_75_cast_fp16")];
-            tensor<int32, [4]> var_2804 = const()[name = tensor<string, []>("op_2804"), val = tensor<int32, [4]>([1, 8, -1, 188])];
-            tensor<fp16, [1, 8, 376, 188]> matrix_bd_77_cast_fp16 = reshape(shape = var_2804, x = matrix_bd_75_cast_fp16)[name = tensor<string, []>("matrix_bd_77_cast_fp16")];
-            tensor<int32, [4]> var_2808_begin_0 = const()[name = tensor<string, []>("op_2808_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
-            tensor<int32, [4]> var_2808_end_0 = const()[name = tensor<string, []>("op_2808_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
-            tensor<bool, [4]> var_2808_end_mask_0 = const()[name = tensor<string, []>("op_2808_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
-            tensor<fp16, [1, 8, 375, 188]> var_2808_cast_fp16 = slice_by_index(begin = var_2808_begin_0, end = var_2808_end_0, end_mask = var_2808_end_mask_0, x = matrix_bd_77_cast_fp16)[name = tensor<string, []>("op_2808_cast_fp16")];
-            tensor<int32, [4]> var_2809 = const()[name = tensor<string, []>("op_2809"), val = tensor<int32, [4]>([1, 8, 188, 375])];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_79_cast_fp16 = reshape(shape = var_2809, x = var_2808_cast_fp16)[name = tensor<string, []>("matrix_bd_79_cast_fp16")];
-            tensor<int32, [4]> var_2814_begin_0 = const()[name = tensor<string, []>("op_2814_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [4]> var_2814_end_0 = const()[name = tensor<string, []>("op_2814_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
-            tensor<bool, [4]> var_2814_end_mask_0 = const()[name = tensor<string, []>("op_2814_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
-            tensor<fp16, [1, 8, 188, 188]> var_2814_cast_fp16 = slice_by_index(begin = var_2814_begin_0, end = var_2814_end_0, end_mask = var_2814_end_mask_0, x = matrix_bd_79_cast_fp16)[name = tensor<string, []>("op_2814_cast_fp16")];
-            tensor<fp16, []> var_2815_to_fp16 = const()[name = tensor<string, []>("op_2815_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 188, 188]> qk_mask_19_cast_fp16 = mul(x = var_2814_cast_fp16, y = var_2815_to_fp16)[name = tensor<string, []>("qk_mask_19_cast_fp16")];
-            tensor<int32, [4]> var_2819 = const()[name = tensor<string, []>("op_2819"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> mh_q_19_cast_fp16 = reshape(shape = var_2819, x = query_39_cast_fp16)[name = tensor<string, []>("mh_q_19_cast_fp16")];
-            tensor<fp16, []> var_2821_to_fp16 = const()[name = tensor<string, []>("op_2821_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 128, 188]> var_2822_cast_fp16 = mul(x = mh_q_19_cast_fp16, y = var_2821_to_fp16)[name = tensor<string, []>("op_2822_cast_fp16")];
-            tensor<int32, [4]> var_2825 = const()[name = tensor<string, []>("op_2825"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_2826_cast_fp16 = reshape(shape = var_2825, x = key_19_cast_fp16)[name = tensor<string, []>("op_2826_cast_fp16")];
+            tensor<int32, [4]> var_2802 = const()[name = tensor<string, []>("op_2802"), val = tensor<int32, [4]>([1, 8, -1, 188])];
+            tensor<fp16, [1, 8, 376, 188]> matrix_bd_77_cast_fp16 = reshape(shape = var_2802, x = matrix_bd_75_cast_fp16)[name = tensor<string, []>("matrix_bd_77_cast_fp16")];
+            tensor<int32, [4]> var_2806_begin_0 = const()[name = tensor<string, []>("op_2806_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2806_end_0 = const()[name = tensor<string, []>("op_2806_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
+            tensor<bool, [4]> var_2806_end_mask_0 = const()[name = tensor<string, []>("op_2806_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 375, 188]> var_2806_cast_fp16 = slice_by_index(begin = var_2806_begin_0, end = var_2806_end_0, end_mask = var_2806_end_mask_0, x = matrix_bd_77_cast_fp16)[name = tensor<string, []>("op_2806_cast_fp16")];
+            tensor<int32, [4]> var_2807 = const()[name = tensor<string, []>("op_2807"), val = tensor<int32, [4]>([1, 8, 188, 375])];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_79_cast_fp16 = reshape(shape = var_2807, x = var_2806_cast_fp16)[name = tensor<string, []>("matrix_bd_79_cast_fp16")];
+            tensor<int32, [4]> var_2812_begin_0 = const()[name = tensor<string, []>("op_2812_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_2812_end_0 = const()[name = tensor<string, []>("op_2812_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
+            tensor<bool, [4]> var_2812_end_mask_0 = const()[name = tensor<string, []>("op_2812_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 188, 188]> var_2812_cast_fp16 = slice_by_index(begin = var_2812_begin_0, end = var_2812_end_0, end_mask = var_2812_end_mask_0, x = matrix_bd_79_cast_fp16)[name = tensor<string, []>("op_2812_cast_fp16")];
+            tensor<fp16, []> var_2813_to_fp16 = const()[name = tensor<string, []>("op_2813_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 188, 188]> qk_mask_19_cast_fp16 = mul(x = var_2812_cast_fp16, y = var_2813_to_fp16)[name = tensor<string, []>("qk_mask_19_cast_fp16")];
+            tensor<int32, [4]> var_2817 = const()[name = tensor<string, []>("op_2817"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> mh_q_19_cast_fp16 = reshape(shape = var_2817, x = query_39_cast_fp16)[name = tensor<string, []>("mh_q_19_cast_fp16")];
+            tensor<fp16, []> var_2819_to_fp16 = const()[name = tensor<string, []>("op_2819_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 128, 188]> var_2820_cast_fp16 = mul(x = mh_q_19_cast_fp16, y = var_2819_to_fp16)[name = tensor<string, []>("op_2820_cast_fp16")];
+            tensor<int32, [4]> var_2823 = const()[name = tensor<string, []>("op_2823"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_2824_cast_fp16 = reshape(shape = var_2823, x = key_19_cast_fp16)[name = tensor<string, []>("op_2824_cast_fp16")];
             tensor<bool, []> mh_w_37_transpose_x_0 = const()[name = tensor<string, []>("mh_w_37_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> mh_w_37_transpose_y_0 = const()[name = tensor<string, []>("mh_w_37_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 188]> mh_w_37_cast_fp16 = matmul(transpose_x = mh_w_37_transpose_x_0, transpose_y = mh_w_37_transpose_y_0, x = var_2822_cast_fp16, y = var_2826_cast_fp16)[name = tensor<string, []>("mh_w_37_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> mh_w_37_cast_fp16 = matmul(transpose_x = mh_w_37_transpose_x_0, transpose_y = mh_w_37_transpose_y_0, x = var_2820_cast_fp16, y = var_2824_cast_fp16)[name = tensor<string, []>("mh_w_37_cast_fp16")];
             tensor<fp16, [1, 8, 188, 188]> mh_w_39_cast_fp16 = add(x = mh_w_37_cast_fp16, y = qk_mask_19_cast_fp16)[name = tensor<string, []>("mh_w_39_cast_fp16")];
-            tensor<fp16, [1, 8, 188, 188]> var_2830_cast_fp16 = softmax(axis = var_2671, x = mh_w_39_cast_fp16)[name = tensor<string, []>("op_2830_cast_fp16")];
-            tensor<int32, [4]> var_2831 = const()[name = tensor<string, []>("op_2831"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_2832_cast_fp16 = reshape(shape = var_2831, x = value_19_cast_fp16)[name = tensor<string, []>("op_2832_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> var_2828_cast_fp16 = softmax(axis = var_2669, x = mh_w_39_cast_fp16)[name = tensor<string, []>("op_2828_cast_fp16")];
+            tensor<int32, [4]> var_2829 = const()[name = tensor<string, []>("op_2829"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_2830_cast_fp16 = reshape(shape = var_2829, x = value_19_cast_fp16)[name = tensor<string, []>("op_2830_cast_fp16")];
             tensor<bool, []> attn_19_transpose_x_0 = const()[name = tensor<string, []>("attn_19_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> attn_19_transpose_y_0 = const()[name = tensor<string, []>("attn_19_transpose_y_0"), val = tensor<bool, []>(true)];
-            tensor<fp16, [1, 8, 128, 188]> attn_19_cast_fp16 = matmul(transpose_x = attn_19_transpose_x_0, transpose_y = attn_19_transpose_y_0, x = var_2832_cast_fp16, y = var_2830_cast_fp16)[name = tensor<string, []>("attn_19_cast_fp16")];
-            tensor<int32, [4]> var_2835 = const()[name = tensor<string, []>("op_2835"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
-            tensor<fp16, [1, 1024, 1, 188]> input_255_cast_fp16 = reshape(shape = var_2835, x = attn_19_cast_fp16)[name = tensor<string, []>("input_255_cast_fp16")];
+            tensor<fp16, [1, 8, 128, 188]> attn_19_cast_fp16 = matmul(transpose_x = attn_19_transpose_x_0, transpose_y = attn_19_transpose_y_0, x = var_2830_cast_fp16, y = var_2828_cast_fp16)[name = tensor<string, []>("attn_19_cast_fp16")];
+            tensor<int32, [4]> var_2833 = const()[name = tensor<string, []>("op_2833"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
+            tensor<fp16, [1, 1024, 1, 188]> input_257_cast_fp16 = reshape(shape = var_2833, x = attn_19_cast_fp16)[name = tensor<string, []>("input_257_cast_fp16")];
             tensor<string, []> obj_41_pad_type_0 = const()[name = tensor<string, []>("obj_41_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> obj_41_strides_0 = const()[name = tensor<string, []>("obj_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> obj_41_pad_0 = const()[name = tensor<string, []>("obj_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> obj_41_dilations_0 = const()[name = tensor<string, []>("obj_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> obj_41_groups_0 = const()[name = tensor<string, []>("obj_41_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_9_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(488034176)))];
-            tensor<fp16, [1, 1024, 1, 188]> obj_41_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = obj_41_dilations_0, groups = obj_41_groups_0, pad = obj_41_pad_0, pad_type = obj_41_pad_type_0, strides = obj_41_strides_0, weight = layers_9_self_attn_o_proj_weight_to_fp16, x = input_255_cast_fp16)[name = tensor<string, []>("obj_41_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_41_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_41_dilations_0, groups = obj_41_groups_0, pad = obj_41_pad_0, pad_type = obj_41_pad_type_0, strides = obj_41_strides_0, weight = layers_9_self_attn_o_proj_weight_to_fp16, x = input_257_cast_fp16)[name = tensor<string, []>("obj_41_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_95_cast_fp16 = add(x = inputs_93_cast_fp16, y = obj_41_cast_fp16)[name = tensor<string, []>("inputs_95_cast_fp16")];
             tensor<int32, [1]> out_95_axes_0 = const()[name = tensor<string, []>("out_95_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_2853_to_fp16 = const()[name = tensor<string, []>("op_2853_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_95_cast_fp16 = layer_norm(axes = out_95_axes_0, epsilon = var_2853_to_fp16, x = inputs_95_cast_fp16)[name = tensor<string, []>("out_95_cast_fp16")];
-            tensor<fp16, [1024]> input_257_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_257_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(490131392)))];
-            tensor<fp16, [1024]> input_257_beta_0_to_fp16 = const()[name = tensor<string, []>("input_257_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(490133504)))];
-            tensor<fp16, []> input_257_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_257_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_257_cast_fp16 = batch_norm(beta = input_257_beta_0_to_fp16, epsilon = input_257_epsilon_0_to_fp16, gamma = input_257_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_95_cast_fp16)[name = tensor<string, []>("input_257_cast_fp16")];
-            tensor<string, []> input_259_pad_type_0 = const()[name = tensor<string, []>("input_259_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_259_strides_0 = const()[name = tensor<string, []>("input_259_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_259_pad_0 = const()[name = tensor<string, []>("input_259_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_259_dilations_0 = const()[name = tensor<string, []>("input_259_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_259_groups_0 = const()[name = tensor<string, []>("input_259_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_2851_to_fp16 = const()[name = tensor<string, []>("op_2851_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_95_cast_fp16 = layer_norm(axes = out_95_axes_0, epsilon = var_2851_to_fp16, x = inputs_95_cast_fp16)[name = tensor<string, []>("out_95_cast_fp16")];
+            tensor<fp16, [1024]> input_259_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_259_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(490131392)))];
+            tensor<fp16, [1024]> input_259_beta_0_to_fp16 = const()[name = tensor<string, []>("input_259_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(490133504)))];
+            tensor<fp16, []> input_259_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_259_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_259_cast_fp16 = batch_norm(beta = input_259_beta_0_to_fp16, epsilon = input_259_epsilon_0_to_fp16, gamma = input_259_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_95_cast_fp16)[name = tensor<string, []>("input_259_cast_fp16")];
+            tensor<string, []> input_261_pad_type_0 = const()[name = tensor<string, []>("input_261_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_261_strides_0 = const()[name = tensor<string, []>("input_261_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_261_pad_0 = const()[name = tensor<string, []>("input_261_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_261_dilations_0 = const()[name = tensor<string, []>("input_261_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_261_groups_0 = const()[name = tensor<string, []>("input_261_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [2048, 1024, 1, 1]> layers_9_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(490135616)))];
-            tensor<fp16, [1, 2048, 1, 188]> input_259_cast_fp16 = conv(dilations = input_259_dilations_0, groups = input_259_groups_0, pad = input_259_pad_0, pad_type = input_259_pad_type_0, strides = input_259_strides_0, weight = layers_9_conv_pointwise_conv1_weight_to_fp16, x = input_257_cast_fp16)[name = tensor<string, []>("input_259_cast_fp16")];
-            tensor<int32, []> input_261_split_num_splits_0 = const()[name = tensor<string, []>("input_261_split_num_splits_0"), val = tensor<int32, []>(2)];
-            tensor<int32, []> input_261_split_axis_0 = const()[name = tensor<string, []>("input_261_split_axis_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1, 1024, 1, 188]> input_261_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_261_split_cast_fp16_1 = split(axis = input_261_split_axis_0, num_splits = input_261_split_num_splits_0, x = input_259_cast_fp16)[name = tensor<string, []>("input_261_split_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_261_split_1_sigmoid_cast_fp16 = sigmoid(x = input_261_split_cast_fp16_1)[name = tensor<string, []>("input_261_split_1_sigmoid_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_261_cast_fp16 = mul(x = input_261_split_cast_fp16_0, y = input_261_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_261_cast_fp16")];
-            tensor<string, []> input_263_pad_type_0 = const()[name = tensor<string, []>("input_263_pad_type_0"), val = tensor<string, []>("custom")];
-            tensor<int32, [4]> input_263_pad_0 = const()[name = tensor<string, []>("input_263_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
-            tensor<int32, []> input_263_groups_0 = const()[name = tensor<string, []>("input_263_groups_0"), val = tensor<int32, []>(1024)];
-            tensor<int32, [2]> input_263_strides_0 = const()[name = tensor<string, []>("input_263_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> input_263_dilations_0 = const()[name = tensor<string, []>("input_263_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [1, 2048, 1, 188]> input_261_cast_fp16 = conv(dilations = input_261_dilations_0, groups = input_261_groups_0, pad = input_261_pad_0, pad_type = input_261_pad_type_0, strides = input_261_strides_0, weight = layers_9_conv_pointwise_conv1_weight_to_fp16, x = input_259_cast_fp16)[name = tensor<string, []>("input_261_cast_fp16")];
+            tensor<int32, []> input_263_split_num_splits_0 = const()[name = tensor<string, []>("input_263_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> input_263_split_axis_0 = const()[name = tensor<string, []>("input_263_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 1, 188]> input_263_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_263_split_cast_fp16_1 = split(axis = input_263_split_axis_0, num_splits = input_263_split_num_splits_0, x = input_261_cast_fp16)[name = tensor<string, []>("input_263_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_263_split_1_sigmoid_cast_fp16 = sigmoid(x = input_263_split_cast_fp16_1)[name = tensor<string, []>("input_263_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_263_cast_fp16 = mul(x = input_263_split_cast_fp16_0, y = input_263_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_263_cast_fp16")];
+            tensor<string, []> input_265_pad_type_0 = const()[name = tensor<string, []>("input_265_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_265_pad_0 = const()[name = tensor<string, []>("input_265_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
+            tensor<int32, []> input_265_groups_0 = const()[name = tensor<string, []>("input_265_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [2]> input_265_strides_0 = const()[name = tensor<string, []>("input_265_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> input_265_dilations_0 = const()[name = tensor<string, []>("input_265_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<fp16, [1024, 1, 1, 9]> const_286_to_fp16 = const()[name = tensor<string, []>("const_286_to_fp16"), val = tensor<fp16, [1024, 1, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(494329984)))];
             tensor<fp16, [1024]> const_287_to_fp16 = const()[name = tensor<string, []>("const_287_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(494348480)))];
-            tensor<fp16, [1, 1024, 1, 188]> input_265_cast_fp16 = conv(bias = const_287_to_fp16, dilations = input_263_dilations_0, groups = input_263_groups_0, pad = input_263_pad_0, pad_type = input_263_pad_type_0, strides = input_263_strides_0, weight = const_286_to_fp16, x = input_261_cast_fp16)[name = tensor<string, []>("input_265_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_267_cast_fp16 = silu(x = input_265_cast_fp16)[name = tensor<string, []>("input_267_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_267_cast_fp16 = conv(bias = const_287_to_fp16, dilations = input_265_dilations_0, groups = input_265_groups_0, pad = input_265_pad_0, pad_type = input_265_pad_type_0, strides = input_265_strides_0, weight = const_286_to_fp16, x = input_263_cast_fp16)[name = tensor<string, []>("input_267_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_269_cast_fp16 = silu(x = input_267_cast_fp16)[name = tensor<string, []>("input_269_cast_fp16")];
             tensor<string, []> x_59_pad_type_0 = const()[name = tensor<string, []>("x_59_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_59_strides_0 = const()[name = tensor<string, []>("x_59_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_59_pad_0 = const()[name = tensor<string, []>("x_59_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_59_dilations_0 = const()[name = tensor<string, []>("x_59_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_59_groups_0 = const()[name = tensor<string, []>("x_59_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_9_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(494350592)))];
-            tensor<fp16, [1, 1024, 1, 188]> x_59_cast_fp16 = conv(dilations = x_59_dilations_0, groups = x_59_groups_0, pad = x_59_pad_0, pad_type = x_59_pad_type_0, strides = x_59_strides_0, weight = layers_9_conv_pointwise_conv2_weight_to_fp16, x = input_267_cast_fp16)[name = tensor<string, []>("x_59_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> x_59_cast_fp16 = conv(dilations = x_59_dilations_0, groups = x_59_groups_0, pad = x_59_pad_0, pad_type = x_59_pad_type_0, strides = x_59_strides_0, weight = layers_9_conv_pointwise_conv2_weight_to_fp16, x = input_269_cast_fp16)[name = tensor<string, []>("x_59_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_97_cast_fp16 = add(x = inputs_95_cast_fp16, y = x_59_cast_fp16)[name = tensor<string, []>("inputs_97_cast_fp16")];
             tensor<int32, [1]> out_97_axes_0 = const()[name = tensor<string, []>("out_97_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_2901_to_fp16 = const()[name = tensor<string, []>("op_2901_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_97_cast_fp16 = layer_norm(axes = out_97_axes_0, epsilon = var_2901_to_fp16, x = inputs_97_cast_fp16)[name = tensor<string, []>("out_97_cast_fp16")];
-            tensor<fp16, [1024]> input_269_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_269_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(496447808)))];
-            tensor<fp16, [1024]> input_269_beta_0_to_fp16 = const()[name = tensor<string, []>("input_269_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(496449920)))];
-            tensor<fp16, []> input_269_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_269_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_269_cast_fp16 = batch_norm(beta = input_269_beta_0_to_fp16, epsilon = input_269_epsilon_0_to_fp16, gamma = input_269_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_97_cast_fp16)[name = tensor<string, []>("input_269_cast_fp16")];
-            tensor<string, []> input_271_pad_type_0 = const()[name = tensor<string, []>("input_271_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_271_strides_0 = const()[name = tensor<string, []>("input_271_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_271_pad_0 = const()[name = tensor<string, []>("input_271_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_271_dilations_0 = const()[name = tensor<string, []>("input_271_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_271_groups_0 = const()[name = tensor<string, []>("input_271_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_2899_to_fp16 = const()[name = tensor<string, []>("op_2899_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_97_cast_fp16 = layer_norm(axes = out_97_axes_0, epsilon = var_2899_to_fp16, x = inputs_97_cast_fp16)[name = tensor<string, []>("out_97_cast_fp16")];
+            tensor<fp16, [1024]> input_271_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_271_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(496447808)))];
+            tensor<fp16, [1024]> input_271_beta_0_to_fp16 = const()[name = tensor<string, []>("input_271_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(496449920)))];
+            tensor<fp16, []> input_271_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_271_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_271_cast_fp16 = batch_norm(beta = input_271_beta_0_to_fp16, epsilon = input_271_epsilon_0_to_fp16, gamma = input_271_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_97_cast_fp16)[name = tensor<string, []>("input_271_cast_fp16")];
+            tensor<string, []> input_273_pad_type_0 = const()[name = tensor<string, []>("input_273_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_273_strides_0 = const()[name = tensor<string, []>("input_273_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_273_pad_0 = const()[name = tensor<string, []>("input_273_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_273_dilations_0 = const()[name = tensor<string, []>("input_273_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_273_groups_0 = const()[name = tensor<string, []>("input_273_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_9_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_feed_forward2_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(496452032)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_271_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_271_dilations_0, groups = input_271_groups_0, pad = input_271_pad_0, pad_type = input_271_pad_type_0, strides = input_271_strides_0, weight = layers_9_feed_forward2_fc1_weight_to_fp16, x = input_269_cast_fp16)[name = tensor<string, []>("input_271_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_273_cast_fp16 = silu(x = input_271_cast_fp16)[name = tensor<string, []>("input_273_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_273_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_273_dilations_0, groups = input_273_groups_0, pad = input_273_pad_0, pad_type = input_273_pad_type_0, strides = input_273_strides_0, weight = layers_9_feed_forward2_fc1_weight_to_fp16, x = input_271_cast_fp16)[name = tensor<string, []>("input_273_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_275_cast_fp16 = silu(x = input_273_cast_fp16)[name = tensor<string, []>("input_275_cast_fp16")];
             tensor<string, []> x_61_pad_type_0 = const()[name = tensor<string, []>("x_61_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_61_strides_0 = const()[name = tensor<string, []>("x_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_61_pad_0 = const()[name = tensor<string, []>("x_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_61_dilations_0 = const()[name = tensor<string, []>("x_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_61_groups_0 = const()[name = tensor<string, []>("x_61_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_2929_weight_0_to_fp16 = const()[name = tensor<string, []>("op_2929_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(504840704)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_2929_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_61_dilations_0, groups = x_61_groups_0, pad = x_61_pad_0, pad_type = x_61_pad_type_0, strides = x_61_strides_0, weight = var_2929_weight_0_to_fp16, x = input_273_cast_fp16)[name = tensor<string, []>("op_2929_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_99_cast_fp16 = add(x = inputs_97_cast_fp16, y = var_2929_cast_fp16)[name = tensor<string, []>("inputs_99_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_2927_weight_0_to_fp16 = const()[name = tensor<string, []>("op_2927_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(504840704)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_2927_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_61_dilations_0, groups = x_61_groups_0, pad = x_61_pad_0, pad_type = x_61_pad_type_0, strides = x_61_strides_0, weight = var_2927_weight_0_to_fp16, x = input_275_cast_fp16)[name = tensor<string, []>("op_2927_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_99_cast_fp16 = add(x = inputs_97_cast_fp16, y = var_2927_cast_fp16)[name = tensor<string, []>("inputs_99_cast_fp16")];
             tensor<int32, [1]> out_99_axes_0 = const()[name = tensor<string, []>("out_99_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_2939_to_fp16 = const()[name = tensor<string, []>("op_2939_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_99_cast_fp16 = layer_norm(axes = out_99_axes_0, epsilon = var_2939_to_fp16, x = inputs_99_cast_fp16)[name = tensor<string, []>("out_99_cast_fp16")];
+            tensor<fp16, []> var_2937_to_fp16 = const()[name = tensor<string, []>("op_2937_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_99_cast_fp16 = layer_norm(axes = out_99_axes_0, epsilon = var_2937_to_fp16, x = inputs_99_cast_fp16)[name = tensor<string, []>("out_99_cast_fp16")];
             tensor<fp16, [1024]> inputs_101_gamma_0_to_fp16 = const()[name = tensor<string, []>("inputs_101_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(513229376)))];
             tensor<fp16, [1024]> inputs_101_beta_0_to_fp16 = const()[name = tensor<string, []>("inputs_101_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(513231488)))];
             tensor<fp16, []> inputs_101_epsilon_0_to_fp16 = const()[name = tensor<string, []>("inputs_101_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_101_cast_fp16 = batch_norm(beta = inputs_101_beta_0_to_fp16, epsilon = inputs_101_epsilon_0_to_fp16, gamma = inputs_101_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_99_cast_fp16)[name = tensor<string, []>("inputs_101_cast_fp16")];
-            tensor<int32, []> var_2953 = const()[name = tensor<string, []>("op_2953"), val = tensor<int32, []>(3)];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_101_cast_fp16 = batch_norm(beta = inputs_101_beta_0_to_fp16, epsilon = inputs_101_epsilon_0_to_fp16, gamma = inputs_101_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_99_cast_fp16)[name = tensor<string, []>("inputs_101_cast_fp16")];
+            tensor<int32, []> var_2951 = const()[name = tensor<string, []>("op_2951"), val = tensor<int32, []>(3)];
             tensor<int32, [1]> out_101_axes_0 = const()[name = tensor<string, []>("out_101_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_2984_to_fp16 = const()[name = tensor<string, []>("op_2984_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_101_cast_fp16 = layer_norm(axes = out_101_axes_0, epsilon = var_2984_to_fp16, x = inputs_101_cast_fp16)[name = tensor<string, []>("out_101_cast_fp16")];
-            tensor<fp16, [1024]> input_275_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_275_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(513233600)))];
-            tensor<fp16, [1024]> input_275_beta_0_to_fp16 = const()[name = tensor<string, []>("input_275_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(513235712)))];
-            tensor<fp16, []> input_275_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_275_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_275_cast_fp16 = batch_norm(beta = input_275_beta_0_to_fp16, epsilon = input_275_epsilon_0_to_fp16, gamma = input_275_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_101_cast_fp16)[name = tensor<string, []>("input_275_cast_fp16")];
-            tensor<string, []> input_277_pad_type_0 = const()[name = tensor<string, []>("input_277_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_277_strides_0 = const()[name = tensor<string, []>("input_277_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_277_pad_0 = const()[name = tensor<string, []>("input_277_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_277_dilations_0 = const()[name = tensor<string, []>("input_277_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_277_groups_0 = const()[name = tensor<string, []>("input_277_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_2982_to_fp16 = const()[name = tensor<string, []>("op_2982_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_101_cast_fp16 = layer_norm(axes = out_101_axes_0, epsilon = var_2982_to_fp16, x = inputs_101_cast_fp16)[name = tensor<string, []>("out_101_cast_fp16")];
+            tensor<fp16, [1024]> input_277_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_277_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(513233600)))];
+            tensor<fp16, [1024]> input_277_beta_0_to_fp16 = const()[name = tensor<string, []>("input_277_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(513235712)))];
+            tensor<fp16, []> input_277_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_277_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_277_cast_fp16 = batch_norm(beta = input_277_beta_0_to_fp16, epsilon = input_277_epsilon_0_to_fp16, gamma = input_277_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_101_cast_fp16)[name = tensor<string, []>("input_277_cast_fp16")];
+            tensor<string, []> input_279_pad_type_0 = const()[name = tensor<string, []>("input_279_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_279_strides_0 = const()[name = tensor<string, []>("input_279_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_279_pad_0 = const()[name = tensor<string, []>("input_279_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_279_dilations_0 = const()[name = tensor<string, []>("input_279_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_279_groups_0 = const()[name = tensor<string, []>("input_279_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_10_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_feed_forward1_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(513237824)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_277_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_277_dilations_0, groups = input_277_groups_0, pad = input_277_pad_0, pad_type = input_277_pad_type_0, strides = input_277_strides_0, weight = layers_10_feed_forward1_fc1_weight_to_fp16, x = input_275_cast_fp16)[name = tensor<string, []>("input_277_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_279_cast_fp16 = silu(x = input_277_cast_fp16)[name = tensor<string, []>("input_279_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_279_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_279_dilations_0, groups = input_279_groups_0, pad = input_279_pad_0, pad_type = input_279_pad_type_0, strides = input_279_strides_0, weight = layers_10_feed_forward1_fc1_weight_to_fp16, x = input_277_cast_fp16)[name = tensor<string, []>("input_279_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_281_cast_fp16 = silu(x = input_279_cast_fp16)[name = tensor<string, []>("input_281_cast_fp16")];
             tensor<string, []> x_63_pad_type_0 = const()[name = tensor<string, []>("x_63_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_63_strides_0 = const()[name = tensor<string, []>("x_63_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_63_pad_0 = const()[name = tensor<string, []>("x_63_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_63_dilations_0 = const()[name = tensor<string, []>("x_63_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_63_groups_0 = const()[name = tensor<string, []>("x_63_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_3012_weight_0_to_fp16 = const()[name = tensor<string, []>("op_3012_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(521626496)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_3012_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_63_dilations_0, groups = x_63_groups_0, pad = x_63_pad_0, pad_type = x_63_pad_type_0, strides = x_63_strides_0, weight = var_3012_weight_0_to_fp16, x = input_279_cast_fp16)[name = tensor<string, []>("op_3012_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_103_cast_fp16 = add(x = inputs_101_cast_fp16, y = var_3012_cast_fp16)[name = tensor<string, []>("inputs_103_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_3010_weight_0_to_fp16 = const()[name = tensor<string, []>("op_3010_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(521626496)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_3010_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_63_dilations_0, groups = x_63_groups_0, pad = x_63_pad_0, pad_type = x_63_pad_type_0, strides = x_63_strides_0, weight = var_3010_weight_0_to_fp16, x = input_281_cast_fp16)[name = tensor<string, []>("op_3010_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_103_cast_fp16 = add(x = inputs_101_cast_fp16, y = var_3010_cast_fp16)[name = tensor<string, []>("inputs_103_cast_fp16")];
             tensor<int32, [1]> out_103_axes_0 = const()[name = tensor<string, []>("out_103_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_3022_to_fp16 = const()[name = tensor<string, []>("op_3022_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_103_cast_fp16 = layer_norm(axes = out_103_axes_0, epsilon = var_3022_to_fp16, x = inputs_103_cast_fp16)[name = tensor<string, []>("out_103_cast_fp16")];
+            tensor<fp16, []> var_3020_to_fp16 = const()[name = tensor<string, []>("op_3020_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_103_cast_fp16 = layer_norm(axes = out_103_axes_0, epsilon = var_3020_to_fp16, x = inputs_103_cast_fp16)[name = tensor<string, []>("out_103_cast_fp16")];
             tensor<fp16, [1024]> obj_43_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_43_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(530015168)))];
             tensor<fp16, [1024]> obj_43_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_43_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(530017280)))];
             tensor<fp16, []> obj_43_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_43_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> obj_43_cast_fp16 = batch_norm(beta = obj_43_beta_0_to_fp16, epsilon = obj_43_epsilon_0_to_fp16, gamma = obj_43_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_103_cast_fp16)[name = tensor<string, []>("obj_43_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_43_cast_fp16 = batch_norm(beta = obj_43_beta_0_to_fp16, epsilon = obj_43_epsilon_0_to_fp16, gamma = obj_43_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_103_cast_fp16)[name = tensor<string, []>("obj_43_cast_fp16")];
             tensor<string, []> query_41_pad_type_0 = const()[name = tensor<string, []>("query_41_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> query_41_strides_0 = const()[name = tensor<string, []>("query_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> query_41_pad_0 = const()[name = tensor<string, []>("query_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> query_41_dilations_0 = const()[name = tensor<string, []>("query_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> query_41_groups_0 = const()[name = tensor<string, []>("query_41_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_10_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(530019392)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_41_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = query_41_dilations_0, groups = query_41_groups_0, pad = query_41_pad_0, pad_type = query_41_pad_type_0, strides = query_41_strides_0, weight = layers_10_self_attn_q_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor<string, []>("query_41_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> query_41_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_41_dilations_0, groups = query_41_groups_0, pad = query_41_pad_0, pad_type = query_41_pad_type_0, strides = query_41_strides_0, weight = layers_10_self_attn_q_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor<string, []>("query_41_cast_fp16")];
             tensor<string, []> key_21_pad_type_0 = const()[name = tensor<string, []>("key_21_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> key_21_strides_0 = const()[name = tensor<string, []>("key_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> key_21_pad_0 = const()[name = tensor<string, []>("key_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
@@ -1912,173 +1914,173 @@ program(1.0)
             tensor<int32, [2]> value_21_dilations_0 = const()[name = tensor<string, []>("value_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> value_21_groups_0 = const()[name = tensor<string, []>("value_21_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_10_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(534213824)))];
-            tensor<fp16, [1, 1024, 1, 188]> value_21_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = value_21_dilations_0, groups = value_21_groups_0, pad = value_21_pad_0, pad_type = value_21_pad_type_0, strides = value_21_strides_0, weight = layers_10_self_attn_v_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor<string, []>("value_21_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_3060_to_fp16 = const()[name = tensor<string, []>("op_3060_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(536311040)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_43_cast_fp16 = add(x = query_41_cast_fp16, y = var_3060_to_fp16)[name = tensor<string, []>("query_43_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_3063_to_fp16 = const()[name = tensor<string, []>("op_3063_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(536313152)))];
-            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_21_cast_fp16 = add(x = query_41_cast_fp16, y = var_3063_to_fp16)[name = tensor<string, []>("q_with_bias_v_21_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> value_21_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_21_dilations_0, groups = value_21_groups_0, pad = value_21_pad_0, pad_type = value_21_pad_type_0, strides = value_21_strides_0, weight = layers_10_self_attn_v_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor<string, []>("value_21_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_3058_to_fp16 = const()[name = tensor<string, []>("op_3058_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(536311040)))];
+            tensor<fp16, [1, 1024, 1, 188]> query_43_cast_fp16 = add(x = query_41_cast_fp16, y = var_3058_to_fp16)[name = tensor<string, []>("query_43_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_3061_to_fp16 = const()[name = tensor<string, []>("op_3061_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(536313152)))];
+            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_21_cast_fp16 = add(x = query_41_cast_fp16, y = var_3061_to_fp16)[name = tensor<string, []>("q_with_bias_v_21_cast_fp16")];
             tensor<string, []> p_21_pad_type_0 = const()[name = tensor<string, []>("p_21_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> p_21_strides_0 = const()[name = tensor<string, []>("p_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> p_21_pad_0 = const()[name = tensor<string, []>("p_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> p_21_dilations_0 = const()[name = tensor<string, []>("p_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> p_21_groups_0 = const()[name = tensor<string, []>("p_21_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_10_self_attn_linear_pos_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_linear_pos_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(536315264)))];
-            tensor<fp16, [1, 1024, 1, 375]> p_21_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = p_21_dilations_0, groups = p_21_groups_0, pad = p_21_pad_0, pad_type = p_21_pad_type_0, strides = p_21_strides_0, weight = layers_10_self_attn_linear_pos_weight_to_fp16, x = pos_enc_to_fp16)[name = tensor<string, []>("p_21_cast_fp16")];
-            tensor<int32, [4]> var_3074 = const()[name = tensor<string, []>("op_3074"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_3075_cast_fp16 = reshape(shape = var_3074, x = q_with_bias_v_21_cast_fp16)[name = tensor<string, []>("op_3075_cast_fp16")];
-            tensor<int32, [4]> var_3076 = const()[name = tensor<string, []>("op_3076"), val = tensor<int32, [4]>([1, 8, 128, -1])];
-            tensor<fp16, [1, 8, 128, 375]> var_3077_cast_fp16 = reshape(shape = var_3076, x = p_21_cast_fp16)[name = tensor<string, []>("op_3077_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 375]> p_21_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_21_dilations_0, groups = p_21_groups_0, pad = p_21_pad_0, pad_type = p_21_pad_type_0, strides = p_21_strides_0, weight = layers_10_self_attn_linear_pos_weight_to_fp16, x = obj_3_cast_fp16)[name = tensor<string, []>("p_21_cast_fp16")];
+            tensor<int32, [4]> var_3072 = const()[name = tensor<string, []>("op_3072"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_3073_cast_fp16 = reshape(shape = var_3072, x = q_with_bias_v_21_cast_fp16)[name = tensor<string, []>("op_3073_cast_fp16")];
+            tensor<int32, [4]> var_3074 = const()[name = tensor<string, []>("op_3074"), val = tensor<int32, [4]>([1, 8, 128, -1])];
+            tensor<fp16, [1, 8, 128, 375]> var_3075_cast_fp16 = reshape(shape = var_3074, x = p_21_cast_fp16)[name = tensor<string, []>("op_3075_cast_fp16")];
             tensor<bool, []> matrix_bd_81_transpose_x_0 = const()[name = tensor<string, []>("matrix_bd_81_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> matrix_bd_81_transpose_y_0 = const()[name = tensor<string, []>("matrix_bd_81_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_81_cast_fp16 = matmul(transpose_x = matrix_bd_81_transpose_x_0, transpose_y = matrix_bd_81_transpose_y_0, x = var_3075_cast_fp16, y = var_3077_cast_fp16)[name = tensor<string, []>("matrix_bd_81_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_81_cast_fp16 = matmul(transpose_x = matrix_bd_81_transpose_x_0, transpose_y = matrix_bd_81_transpose_y_0, x = var_3073_cast_fp16, y = var_3075_cast_fp16)[name = tensor<string, []>("matrix_bd_81_cast_fp16")];
             tensor<int32, [8]> matrix_bd_83_pad_0 = const()[name = tensor<string, []>("matrix_bd_83_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
             tensor<string, []> matrix_bd_83_mode_0 = const()[name = tensor<string, []>("matrix_bd_83_mode_0"), val = tensor<string, []>("constant")];
             tensor<fp16, []> const_120_to_fp16 = const()[name = tensor<string, []>("const_120_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
             tensor<fp16, [1, 8, 188, 376]> matrix_bd_83_cast_fp16 = pad(constant_val = const_120_to_fp16, mode = matrix_bd_83_mode_0, pad = matrix_bd_83_pad_0, x = matrix_bd_81_cast_fp16)[name = tensor<string, []>("matrix_bd_83_cast_fp16")];
-            tensor<int32, [4]> var_3086 = const()[name = tensor<string, []>("op_3086"), val = tensor<int32, [4]>([1, 8, -1, 188])];
-            tensor<fp16, [1, 8, 376, 188]> matrix_bd_85_cast_fp16 = reshape(shape = var_3086, x = matrix_bd_83_cast_fp16)[name = tensor<string, []>("matrix_bd_85_cast_fp16")];
-            tensor<int32, [4]> var_3090_begin_0 = const()[name = tensor<string, []>("op_3090_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
-            tensor<int32, [4]> var_3090_end_0 = const()[name = tensor<string, []>("op_3090_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
-            tensor<bool, [4]> var_3090_end_mask_0 = const()[name = tensor<string, []>("op_3090_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
-            tensor<fp16, [1, 8, 375, 188]> var_3090_cast_fp16 = slice_by_index(begin = var_3090_begin_0, end = var_3090_end_0, end_mask = var_3090_end_mask_0, x = matrix_bd_85_cast_fp16)[name = tensor<string, []>("op_3090_cast_fp16")];
-            tensor<int32, [4]> var_3091 = const()[name = tensor<string, []>("op_3091"), val = tensor<int32, [4]>([1, 8, 188, 375])];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_87_cast_fp16 = reshape(shape = var_3091, x = var_3090_cast_fp16)[name = tensor<string, []>("matrix_bd_87_cast_fp16")];
-            tensor<int32, [4]> var_3096_begin_0 = const()[name = tensor<string, []>("op_3096_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [4]> var_3096_end_0 = const()[name = tensor<string, []>("op_3096_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
-            tensor<bool, [4]> var_3096_end_mask_0 = const()[name = tensor<string, []>("op_3096_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
-            tensor<fp16, [1, 8, 188, 188]> var_3096_cast_fp16 = slice_by_index(begin = var_3096_begin_0, end = var_3096_end_0, end_mask = var_3096_end_mask_0, x = matrix_bd_87_cast_fp16)[name = tensor<string, []>("op_3096_cast_fp16")];
-            tensor<fp16, []> var_3097_to_fp16 = const()[name = tensor<string, []>("op_3097_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 188, 188]> qk_mask_21_cast_fp16 = mul(x = var_3096_cast_fp16, y = var_3097_to_fp16)[name = tensor<string, []>("qk_mask_21_cast_fp16")];
-            tensor<int32, [4]> var_3101 = const()[name = tensor<string, []>("op_3101"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> mh_q_21_cast_fp16 = reshape(shape = var_3101, x = query_43_cast_fp16)[name = tensor<string, []>("mh_q_21_cast_fp16")];
-            tensor<fp16, []> var_3103_to_fp16 = const()[name = tensor<string, []>("op_3103_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 128, 188]> var_3104_cast_fp16 = mul(x = mh_q_21_cast_fp16, y = var_3103_to_fp16)[name = tensor<string, []>("op_3104_cast_fp16")];
-            tensor<int32, [4]> var_3107 = const()[name = tensor<string, []>("op_3107"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_3108_cast_fp16 = reshape(shape = var_3107, x = key_21_cast_fp16)[name = tensor<string, []>("op_3108_cast_fp16")];
+            tensor<int32, [4]> var_3084 = const()[name = tensor<string, []>("op_3084"), val = tensor<int32, [4]>([1, 8, -1, 188])];
+            tensor<fp16, [1, 8, 376, 188]> matrix_bd_85_cast_fp16 = reshape(shape = var_3084, x = matrix_bd_83_cast_fp16)[name = tensor<string, []>("matrix_bd_85_cast_fp16")];
+            tensor<int32, [4]> var_3088_begin_0 = const()[name = tensor<string, []>("op_3088_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3088_end_0 = const()[name = tensor<string, []>("op_3088_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
+            tensor<bool, [4]> var_3088_end_mask_0 = const()[name = tensor<string, []>("op_3088_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 375, 188]> var_3088_cast_fp16 = slice_by_index(begin = var_3088_begin_0, end = var_3088_end_0, end_mask = var_3088_end_mask_0, x = matrix_bd_85_cast_fp16)[name = tensor<string, []>("op_3088_cast_fp16")];
+            tensor<int32, [4]> var_3089 = const()[name = tensor<string, []>("op_3089"), val = tensor<int32, [4]>([1, 8, 188, 375])];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_87_cast_fp16 = reshape(shape = var_3089, x = var_3088_cast_fp16)[name = tensor<string, []>("matrix_bd_87_cast_fp16")];
+            tensor<int32, [4]> var_3094_begin_0 = const()[name = tensor<string, []>("op_3094_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_3094_end_0 = const()[name = tensor<string, []>("op_3094_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
+            tensor<bool, [4]> var_3094_end_mask_0 = const()[name = tensor<string, []>("op_3094_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 188, 188]> var_3094_cast_fp16 = slice_by_index(begin = var_3094_begin_0, end = var_3094_end_0, end_mask = var_3094_end_mask_0, x = matrix_bd_87_cast_fp16)[name = tensor<string, []>("op_3094_cast_fp16")];
+            tensor<fp16, []> var_3095_to_fp16 = const()[name = tensor<string, []>("op_3095_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 188, 188]> qk_mask_21_cast_fp16 = mul(x = var_3094_cast_fp16, y = var_3095_to_fp16)[name = tensor<string, []>("qk_mask_21_cast_fp16")];
+            tensor<int32, [4]> var_3099 = const()[name = tensor<string, []>("op_3099"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> mh_q_21_cast_fp16 = reshape(shape = var_3099, x = query_43_cast_fp16)[name = tensor<string, []>("mh_q_21_cast_fp16")];
+            tensor<fp16, []> var_3101_to_fp16 = const()[name = tensor<string, []>("op_3101_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 128, 188]> var_3102_cast_fp16 = mul(x = mh_q_21_cast_fp16, y = var_3101_to_fp16)[name = tensor<string, []>("op_3102_cast_fp16")];
+            tensor<int32, [4]> var_3105 = const()[name = tensor<string, []>("op_3105"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_3106_cast_fp16 = reshape(shape = var_3105, x = key_21_cast_fp16)[name = tensor<string, []>("op_3106_cast_fp16")];
             tensor<bool, []> mh_w_41_transpose_x_0 = const()[name = tensor<string, []>("mh_w_41_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> mh_w_41_transpose_y_0 = const()[name = tensor<string, []>("mh_w_41_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 188]> mh_w_41_cast_fp16 = matmul(transpose_x = mh_w_41_transpose_x_0, transpose_y = mh_w_41_transpose_y_0, x = var_3104_cast_fp16, y = var_3108_cast_fp16)[name = tensor<string, []>("mh_w_41_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> mh_w_41_cast_fp16 = matmul(transpose_x = mh_w_41_transpose_x_0, transpose_y = mh_w_41_transpose_y_0, x = var_3102_cast_fp16, y = var_3106_cast_fp16)[name = tensor<string, []>("mh_w_41_cast_fp16")];
             tensor<fp16, [1, 8, 188, 188]> mh_w_43_cast_fp16 = add(x = mh_w_41_cast_fp16, y = qk_mask_21_cast_fp16)[name = tensor<string, []>("mh_w_43_cast_fp16")];
-            tensor<fp16, [1, 8, 188, 188]> var_3112_cast_fp16 = softmax(axis = var_2953, x = mh_w_43_cast_fp16)[name = tensor<string, []>("op_3112_cast_fp16")];
-            tensor<int32, [4]> var_3113 = const()[name = tensor<string, []>("op_3113"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_3114_cast_fp16 = reshape(shape = var_3113, x = value_21_cast_fp16)[name = tensor<string, []>("op_3114_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> var_3110_cast_fp16 = softmax(axis = var_2951, x = mh_w_43_cast_fp16)[name = tensor<string, []>("op_3110_cast_fp16")];
+            tensor<int32, [4]> var_3111 = const()[name = tensor<string, []>("op_3111"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_3112_cast_fp16 = reshape(shape = var_3111, x = value_21_cast_fp16)[name = tensor<string, []>("op_3112_cast_fp16")];
             tensor<bool, []> attn_21_transpose_x_0 = const()[name = tensor<string, []>("attn_21_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> attn_21_transpose_y_0 = const()[name = tensor<string, []>("attn_21_transpose_y_0"), val = tensor<bool, []>(true)];
-            tensor<fp16, [1, 8, 128, 188]> attn_21_cast_fp16 = matmul(transpose_x = attn_21_transpose_x_0, transpose_y = attn_21_transpose_y_0, x = var_3114_cast_fp16, y = var_3112_cast_fp16)[name = tensor<string, []>("attn_21_cast_fp16")];
-            tensor<int32, [4]> var_3117 = const()[name = tensor<string, []>("op_3117"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
-            tensor<fp16, [1, 1024, 1, 188]> input_281_cast_fp16 = reshape(shape = var_3117, x = attn_21_cast_fp16)[name = tensor<string, []>("input_281_cast_fp16")];
+            tensor<fp16, [1, 8, 128, 188]> attn_21_cast_fp16 = matmul(transpose_x = attn_21_transpose_x_0, transpose_y = attn_21_transpose_y_0, x = var_3112_cast_fp16, y = var_3110_cast_fp16)[name = tensor<string, []>("attn_21_cast_fp16")];
+            tensor<int32, [4]> var_3115 = const()[name = tensor<string, []>("op_3115"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
+            tensor<fp16, [1, 1024, 1, 188]> input_283_cast_fp16 = reshape(shape = var_3115, x = attn_21_cast_fp16)[name = tensor<string, []>("input_283_cast_fp16")];
             tensor<string, []> obj_45_pad_type_0 = const()[name = tensor<string, []>("obj_45_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> obj_45_strides_0 = const()[name = tensor<string, []>("obj_45_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> obj_45_pad_0 = const()[name = tensor<string, []>("obj_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> obj_45_dilations_0 = const()[name = tensor<string, []>("obj_45_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> obj_45_groups_0 = const()[name = tensor<string, []>("obj_45_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_10_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(538412480)))];
-            tensor<fp16, [1, 1024, 1, 188]> obj_45_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = obj_45_dilations_0, groups = obj_45_groups_0, pad = obj_45_pad_0, pad_type = obj_45_pad_type_0, strides = obj_45_strides_0, weight = layers_10_self_attn_o_proj_weight_to_fp16, x = input_281_cast_fp16)[name = tensor<string, []>("obj_45_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_45_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_45_dilations_0, groups = obj_45_groups_0, pad = obj_45_pad_0, pad_type = obj_45_pad_type_0, strides = obj_45_strides_0, weight = layers_10_self_attn_o_proj_weight_to_fp16, x = input_283_cast_fp16)[name = tensor<string, []>("obj_45_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_105_cast_fp16 = add(x = inputs_103_cast_fp16, y = obj_45_cast_fp16)[name = tensor<string, []>("inputs_105_cast_fp16")];
             tensor<int32, [1]> out_105_axes_0 = const()[name = tensor<string, []>("out_105_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_3135_to_fp16 = const()[name = tensor<string, []>("op_3135_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_105_cast_fp16 = layer_norm(axes = out_105_axes_0, epsilon = var_3135_to_fp16, x = inputs_105_cast_fp16)[name = tensor<string, []>("out_105_cast_fp16")];
-            tensor<fp16, [1024]> input_283_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_283_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(540509696)))];
-            tensor<fp16, [1024]> input_283_beta_0_to_fp16 = const()[name = tensor<string, []>("input_283_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(540511808)))];
-            tensor<fp16, []> input_283_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_283_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_283_cast_fp16 = batch_norm(beta = input_283_beta_0_to_fp16, epsilon = input_283_epsilon_0_to_fp16, gamma = input_283_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_105_cast_fp16)[name = tensor<string, []>("input_283_cast_fp16")];
-            tensor<string, []> input_285_pad_type_0 = const()[name = tensor<string, []>("input_285_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_285_strides_0 = const()[name = tensor<string, []>("input_285_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_285_pad_0 = const()[name = tensor<string, []>("input_285_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_285_dilations_0 = const()[name = tensor<string, []>("input_285_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_285_groups_0 = const()[name = tensor<string, []>("input_285_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_3133_to_fp16 = const()[name = tensor<string, []>("op_3133_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_105_cast_fp16 = layer_norm(axes = out_105_axes_0, epsilon = var_3133_to_fp16, x = inputs_105_cast_fp16)[name = tensor<string, []>("out_105_cast_fp16")];
+            tensor<fp16, [1024]> input_285_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_285_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(540509696)))];
+            tensor<fp16, [1024]> input_285_beta_0_to_fp16 = const()[name = tensor<string, []>("input_285_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(540511808)))];
+            tensor<fp16, []> input_285_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_285_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_285_cast_fp16 = batch_norm(beta = input_285_beta_0_to_fp16, epsilon = input_285_epsilon_0_to_fp16, gamma = input_285_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_105_cast_fp16)[name = tensor<string, []>("input_285_cast_fp16")];
+            tensor<string, []> input_287_pad_type_0 = const()[name = tensor<string, []>("input_287_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_287_strides_0 = const()[name = tensor<string, []>("input_287_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_287_pad_0 = const()[name = tensor<string, []>("input_287_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_287_dilations_0 = const()[name = tensor<string, []>("input_287_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_287_groups_0 = const()[name = tensor<string, []>("input_287_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [2048, 1024, 1, 1]> layers_10_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(540513920)))];
-            tensor<fp16, [1, 2048, 1, 188]> input_285_cast_fp16 = conv(dilations = input_285_dilations_0, groups = input_285_groups_0, pad = input_285_pad_0, pad_type = input_285_pad_type_0, strides = input_285_strides_0, weight = layers_10_conv_pointwise_conv1_weight_to_fp16, x = input_283_cast_fp16)[name = tensor<string, []>("input_285_cast_fp16")];
-            tensor<int32, []> input_287_split_num_splits_0 = const()[name = tensor<string, []>("input_287_split_num_splits_0"), val = tensor<int32, []>(2)];
-            tensor<int32, []> input_287_split_axis_0 = const()[name = tensor<string, []>("input_287_split_axis_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1, 1024, 1, 188]> input_287_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_287_split_cast_fp16_1 = split(axis = input_287_split_axis_0, num_splits = input_287_split_num_splits_0, x = input_285_cast_fp16)[name = tensor<string, []>("input_287_split_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_287_split_1_sigmoid_cast_fp16 = sigmoid(x = input_287_split_cast_fp16_1)[name = tensor<string, []>("input_287_split_1_sigmoid_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_287_cast_fp16 = mul(x = input_287_split_cast_fp16_0, y = input_287_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_287_cast_fp16")];
-            tensor<string, []> input_289_pad_type_0 = const()[name = tensor<string, []>("input_289_pad_type_0"), val = tensor<string, []>("custom")];
-            tensor<int32, [4]> input_289_pad_0 = const()[name = tensor<string, []>("input_289_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
-            tensor<int32, []> input_289_groups_0 = const()[name = tensor<string, []>("input_289_groups_0"), val = tensor<int32, []>(1024)];
-            tensor<int32, [2]> input_289_strides_0 = const()[name = tensor<string, []>("input_289_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> input_289_dilations_0 = const()[name = tensor<string, []>("input_289_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [1, 2048, 1, 188]> input_287_cast_fp16 = conv(dilations = input_287_dilations_0, groups = input_287_groups_0, pad = input_287_pad_0, pad_type = input_287_pad_type_0, strides = input_287_strides_0, weight = layers_10_conv_pointwise_conv1_weight_to_fp16, x = input_285_cast_fp16)[name = tensor<string, []>("input_287_cast_fp16")];
+            tensor<int32, []> input_289_split_num_splits_0 = const()[name = tensor<string, []>("input_289_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> input_289_split_axis_0 = const()[name = tensor<string, []>("input_289_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 1, 188]> input_289_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_289_split_cast_fp16_1 = split(axis = input_289_split_axis_0, num_splits = input_289_split_num_splits_0, x = input_287_cast_fp16)[name = tensor<string, []>("input_289_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_289_split_1_sigmoid_cast_fp16 = sigmoid(x = input_289_split_cast_fp16_1)[name = tensor<string, []>("input_289_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_289_cast_fp16 = mul(x = input_289_split_cast_fp16_0, y = input_289_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_289_cast_fp16")];
+            tensor<string, []> input_291_pad_type_0 = const()[name = tensor<string, []>("input_291_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_291_pad_0 = const()[name = tensor<string, []>("input_291_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
+            tensor<int32, []> input_291_groups_0 = const()[name = tensor<string, []>("input_291_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [2]> input_291_strides_0 = const()[name = tensor<string, []>("input_291_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> input_291_dilations_0 = const()[name = tensor<string, []>("input_291_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<fp16, [1024, 1, 1, 9]> const_288_to_fp16 = const()[name = tensor<string, []>("const_288_to_fp16"), val = tensor<fp16, [1024, 1, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(544708288)))];
             tensor<fp16, [1024]> const_289_to_fp16 = const()[name = tensor<string, []>("const_289_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(544726784)))];
-            tensor<fp16, [1, 1024, 1, 188]> input_291_cast_fp16 = conv(bias = const_289_to_fp16, dilations = input_289_dilations_0, groups = input_289_groups_0, pad = input_289_pad_0, pad_type = input_289_pad_type_0, strides = input_289_strides_0, weight = const_288_to_fp16, x = input_287_cast_fp16)[name = tensor<string, []>("input_291_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_293_cast_fp16 = silu(x = input_291_cast_fp16)[name = tensor<string, []>("input_293_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_293_cast_fp16 = conv(bias = const_289_to_fp16, dilations = input_291_dilations_0, groups = input_291_groups_0, pad = input_291_pad_0, pad_type = input_291_pad_type_0, strides = input_291_strides_0, weight = const_288_to_fp16, x = input_289_cast_fp16)[name = tensor<string, []>("input_293_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_295_cast_fp16 = silu(x = input_293_cast_fp16)[name = tensor<string, []>("input_295_cast_fp16")];
             tensor<string, []> x_65_pad_type_0 = const()[name = tensor<string, []>("x_65_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_65_strides_0 = const()[name = tensor<string, []>("x_65_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_65_pad_0 = const()[name = tensor<string, []>("x_65_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_65_dilations_0 = const()[name = tensor<string, []>("x_65_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_65_groups_0 = const()[name = tensor<string, []>("x_65_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_10_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(544728896)))];
-            tensor<fp16, [1, 1024, 1, 188]> x_65_cast_fp16 = conv(dilations = x_65_dilations_0, groups = x_65_groups_0, pad = x_65_pad_0, pad_type = x_65_pad_type_0, strides = x_65_strides_0, weight = layers_10_conv_pointwise_conv2_weight_to_fp16, x = input_293_cast_fp16)[name = tensor<string, []>("x_65_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> x_65_cast_fp16 = conv(dilations = x_65_dilations_0, groups = x_65_groups_0, pad = x_65_pad_0, pad_type = x_65_pad_type_0, strides = x_65_strides_0, weight = layers_10_conv_pointwise_conv2_weight_to_fp16, x = input_295_cast_fp16)[name = tensor<string, []>("x_65_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_107_cast_fp16 = add(x = inputs_105_cast_fp16, y = x_65_cast_fp16)[name = tensor<string, []>("inputs_107_cast_fp16")];
             tensor<int32, [1]> out_107_axes_0 = const()[name = tensor<string, []>("out_107_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_3183_to_fp16 = const()[name = tensor<string, []>("op_3183_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_107_cast_fp16 = layer_norm(axes = out_107_axes_0, epsilon = var_3183_to_fp16, x = inputs_107_cast_fp16)[name = tensor<string, []>("out_107_cast_fp16")];
-            tensor<fp16, [1024]> input_295_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_295_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(546826112)))];
-            tensor<fp16, [1024]> input_295_beta_0_to_fp16 = const()[name = tensor<string, []>("input_295_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(546828224)))];
-            tensor<fp16, []> input_295_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_295_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_295_cast_fp16 = batch_norm(beta = input_295_beta_0_to_fp16, epsilon = input_295_epsilon_0_to_fp16, gamma = input_295_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_107_cast_fp16)[name = tensor<string, []>("input_295_cast_fp16")];
-            tensor<string, []> input_297_pad_type_0 = const()[name = tensor<string, []>("input_297_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_297_strides_0 = const()[name = tensor<string, []>("input_297_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_297_pad_0 = const()[name = tensor<string, []>("input_297_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_297_dilations_0 = const()[name = tensor<string, []>("input_297_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_297_groups_0 = const()[name = tensor<string, []>("input_297_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_3181_to_fp16 = const()[name = tensor<string, []>("op_3181_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_107_cast_fp16 = layer_norm(axes = out_107_axes_0, epsilon = var_3181_to_fp16, x = inputs_107_cast_fp16)[name = tensor<string, []>("out_107_cast_fp16")];
+            tensor<fp16, [1024]> input_297_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_297_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(546826112)))];
+            tensor<fp16, [1024]> input_297_beta_0_to_fp16 = const()[name = tensor<string, []>("input_297_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(546828224)))];
+            tensor<fp16, []> input_297_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_297_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_297_cast_fp16 = batch_norm(beta = input_297_beta_0_to_fp16, epsilon = input_297_epsilon_0_to_fp16, gamma = input_297_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_107_cast_fp16)[name = tensor<string, []>("input_297_cast_fp16")];
+            tensor<string, []> input_299_pad_type_0 = const()[name = tensor<string, []>("input_299_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_299_strides_0 = const()[name = tensor<string, []>("input_299_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_299_pad_0 = const()[name = tensor<string, []>("input_299_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_299_dilations_0 = const()[name = tensor<string, []>("input_299_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_299_groups_0 = const()[name = tensor<string, []>("input_299_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_10_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_feed_forward2_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(546830336)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_297_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_297_dilations_0, groups = input_297_groups_0, pad = input_297_pad_0, pad_type = input_297_pad_type_0, strides = input_297_strides_0, weight = layers_10_feed_forward2_fc1_weight_to_fp16, x = input_295_cast_fp16)[name = tensor<string, []>("input_297_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_299_cast_fp16 = silu(x = input_297_cast_fp16)[name = tensor<string, []>("input_299_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_299_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_299_dilations_0, groups = input_299_groups_0, pad = input_299_pad_0, pad_type = input_299_pad_type_0, strides = input_299_strides_0, weight = layers_10_feed_forward2_fc1_weight_to_fp16, x = input_297_cast_fp16)[name = tensor<string, []>("input_299_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_301_cast_fp16 = silu(x = input_299_cast_fp16)[name = tensor<string, []>("input_301_cast_fp16")];
             tensor<string, []> x_67_pad_type_0 = const()[name = tensor<string, []>("x_67_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_67_strides_0 = const()[name = tensor<string, []>("x_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_67_pad_0 = const()[name = tensor<string, []>("x_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_67_dilations_0 = const()[name = tensor<string, []>("x_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_67_groups_0 = const()[name = tensor<string, []>("x_67_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_3211_weight_0_to_fp16 = const()[name = tensor<string, []>("op_3211_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(555219008)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_3211_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_67_dilations_0, groups = x_67_groups_0, pad = x_67_pad_0, pad_type = x_67_pad_type_0, strides = x_67_strides_0, weight = var_3211_weight_0_to_fp16, x = input_299_cast_fp16)[name = tensor<string, []>("op_3211_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_109_cast_fp16 = add(x = inputs_107_cast_fp16, y = var_3211_cast_fp16)[name = tensor<string, []>("inputs_109_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_3209_weight_0_to_fp16 = const()[name = tensor<string, []>("op_3209_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(555219008)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_3209_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_67_dilations_0, groups = x_67_groups_0, pad = x_67_pad_0, pad_type = x_67_pad_type_0, strides = x_67_strides_0, weight = var_3209_weight_0_to_fp16, x = input_301_cast_fp16)[name = tensor<string, []>("op_3209_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_109_cast_fp16 = add(x = inputs_107_cast_fp16, y = var_3209_cast_fp16)[name = tensor<string, []>("inputs_109_cast_fp16")];
             tensor<int32, [1]> out_109_axes_0 = const()[name = tensor<string, []>("out_109_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_3221_to_fp16 = const()[name = tensor<string, []>("op_3221_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_109_cast_fp16 = layer_norm(axes = out_109_axes_0, epsilon = var_3221_to_fp16, x = inputs_109_cast_fp16)[name = tensor<string, []>("out_109_cast_fp16")];
+            tensor<fp16, []> var_3219_to_fp16 = const()[name = tensor<string, []>("op_3219_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_109_cast_fp16 = layer_norm(axes = out_109_axes_0, epsilon = var_3219_to_fp16, x = inputs_109_cast_fp16)[name = tensor<string, []>("out_109_cast_fp16")];
             tensor<fp16, [1024]> inputs_111_gamma_0_to_fp16 = const()[name = tensor<string, []>("inputs_111_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(563607680)))];
             tensor<fp16, [1024]> inputs_111_beta_0_to_fp16 = const()[name = tensor<string, []>("inputs_111_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(563609792)))];
             tensor<fp16, []> inputs_111_epsilon_0_to_fp16 = const()[name = tensor<string, []>("inputs_111_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_111_cast_fp16 = batch_norm(beta = inputs_111_beta_0_to_fp16, epsilon = inputs_111_epsilon_0_to_fp16, gamma = inputs_111_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_109_cast_fp16)[name = tensor<string, []>("inputs_111_cast_fp16")];
-            tensor<int32, []> var_3235 = const()[name = tensor<string, []>("op_3235"), val = tensor<int32, []>(3)];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_111_cast_fp16 = batch_norm(beta = inputs_111_beta_0_to_fp16, epsilon = inputs_111_epsilon_0_to_fp16, gamma = inputs_111_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_109_cast_fp16)[name = tensor<string, []>("inputs_111_cast_fp16")];
+            tensor<int32, []> var_3233 = const()[name = tensor<string, []>("op_3233"), val = tensor<int32, []>(3)];
             tensor<int32, [1]> out_111_axes_0 = const()[name = tensor<string, []>("out_111_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_3266_to_fp16 = const()[name = tensor<string, []>("op_3266_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_111_cast_fp16 = layer_norm(axes = out_111_axes_0, epsilon = var_3266_to_fp16, x = inputs_111_cast_fp16)[name = tensor<string, []>("out_111_cast_fp16")];
-            tensor<fp16, [1024]> input_301_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_301_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(563611904)))];
-            tensor<fp16, [1024]> input_301_beta_0_to_fp16 = const()[name = tensor<string, []>("input_301_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(563614016)))];
-            tensor<fp16, []> input_301_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_301_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_301_cast_fp16 = batch_norm(beta = input_301_beta_0_to_fp16, epsilon = input_301_epsilon_0_to_fp16, gamma = input_301_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_111_cast_fp16)[name = tensor<string, []>("input_301_cast_fp16")];
-            tensor<string, []> input_303_pad_type_0 = const()[name = tensor<string, []>("input_303_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_303_strides_0 = const()[name = tensor<string, []>("input_303_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_303_pad_0 = const()[name = tensor<string, []>("input_303_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_303_dilations_0 = const()[name = tensor<string, []>("input_303_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_303_groups_0 = const()[name = tensor<string, []>("input_303_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_3264_to_fp16 = const()[name = tensor<string, []>("op_3264_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_111_cast_fp16 = layer_norm(axes = out_111_axes_0, epsilon = var_3264_to_fp16, x = inputs_111_cast_fp16)[name = tensor<string, []>("out_111_cast_fp16")];
+            tensor<fp16, [1024]> input_303_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_303_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(563611904)))];
+            tensor<fp16, [1024]> input_303_beta_0_to_fp16 = const()[name = tensor<string, []>("input_303_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(563614016)))];
+            tensor<fp16, []> input_303_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_303_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_303_cast_fp16 = batch_norm(beta = input_303_beta_0_to_fp16, epsilon = input_303_epsilon_0_to_fp16, gamma = input_303_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_111_cast_fp16)[name = tensor<string, []>("input_303_cast_fp16")];
+            tensor<string, []> input_305_pad_type_0 = const()[name = tensor<string, []>("input_305_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_305_strides_0 = const()[name = tensor<string, []>("input_305_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_305_pad_0 = const()[name = tensor<string, []>("input_305_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_305_dilations_0 = const()[name = tensor<string, []>("input_305_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_305_groups_0 = const()[name = tensor<string, []>("input_305_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_11_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_feed_forward1_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(563616128)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_303_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_303_dilations_0, groups = input_303_groups_0, pad = input_303_pad_0, pad_type = input_303_pad_type_0, strides = input_303_strides_0, weight = layers_11_feed_forward1_fc1_weight_to_fp16, x = input_301_cast_fp16)[name = tensor<string, []>("input_303_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_305_cast_fp16 = silu(x = input_303_cast_fp16)[name = tensor<string, []>("input_305_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_305_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_305_dilations_0, groups = input_305_groups_0, pad = input_305_pad_0, pad_type = input_305_pad_type_0, strides = input_305_strides_0, weight = layers_11_feed_forward1_fc1_weight_to_fp16, x = input_303_cast_fp16)[name = tensor<string, []>("input_305_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_307_cast_fp16 = silu(x = input_305_cast_fp16)[name = tensor<string, []>("input_307_cast_fp16")];
             tensor<string, []> x_69_pad_type_0 = const()[name = tensor<string, []>("x_69_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_69_strides_0 = const()[name = tensor<string, []>("x_69_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_69_pad_0 = const()[name = tensor<string, []>("x_69_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_69_dilations_0 = const()[name = tensor<string, []>("x_69_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_69_groups_0 = const()[name = tensor<string, []>("x_69_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_3294_weight_0_to_fp16 = const()[name = tensor<string, []>("op_3294_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(572004800)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_3294_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_69_dilations_0, groups = x_69_groups_0, pad = x_69_pad_0, pad_type = x_69_pad_type_0, strides = x_69_strides_0, weight = var_3294_weight_0_to_fp16, x = input_305_cast_fp16)[name = tensor<string, []>("op_3294_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_113_cast_fp16 = add(x = inputs_111_cast_fp16, y = var_3294_cast_fp16)[name = tensor<string, []>("inputs_113_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_3292_weight_0_to_fp16 = const()[name = tensor<string, []>("op_3292_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(572004800)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_3292_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_69_dilations_0, groups = x_69_groups_0, pad = x_69_pad_0, pad_type = x_69_pad_type_0, strides = x_69_strides_0, weight = var_3292_weight_0_to_fp16, x = input_307_cast_fp16)[name = tensor<string, []>("op_3292_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_113_cast_fp16 = add(x = inputs_111_cast_fp16, y = var_3292_cast_fp16)[name = tensor<string, []>("inputs_113_cast_fp16")];
             tensor<int32, [1]> out_113_axes_0 = const()[name = tensor<string, []>("out_113_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_3304_to_fp16 = const()[name = tensor<string, []>("op_3304_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_113_cast_fp16 = layer_norm(axes = out_113_axes_0, epsilon = var_3304_to_fp16, x = inputs_113_cast_fp16)[name = tensor<string, []>("out_113_cast_fp16")];
+            tensor<fp16, []> var_3302_to_fp16 = const()[name = tensor<string, []>("op_3302_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_113_cast_fp16 = layer_norm(axes = out_113_axes_0, epsilon = var_3302_to_fp16, x = inputs_113_cast_fp16)[name = tensor<string, []>("out_113_cast_fp16")];
             tensor<fp16, [1024]> obj_47_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_47_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(580393472)))];
             tensor<fp16, [1024]> obj_47_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_47_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(580395584)))];
             tensor<fp16, []> obj_47_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_47_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> obj_47_cast_fp16 = batch_norm(beta = obj_47_beta_0_to_fp16, epsilon = obj_47_epsilon_0_to_fp16, gamma = obj_47_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_113_cast_fp16)[name = tensor<string, []>("obj_47_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_47_cast_fp16 = batch_norm(beta = obj_47_beta_0_to_fp16, epsilon = obj_47_epsilon_0_to_fp16, gamma = obj_47_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_113_cast_fp16)[name = tensor<string, []>("obj_47_cast_fp16")];
             tensor<string, []> query_45_pad_type_0 = const()[name = tensor<string, []>("query_45_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> query_45_strides_0 = const()[name = tensor<string, []>("query_45_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> query_45_pad_0 = const()[name = tensor<string, []>("query_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> query_45_dilations_0 = const()[name = tensor<string, []>("query_45_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> query_45_groups_0 = const()[name = tensor<string, []>("query_45_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_11_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(580397696)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_45_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = query_45_dilations_0, groups = query_45_groups_0, pad = query_45_pad_0, pad_type = query_45_pad_type_0, strides = query_45_strides_0, weight = layers_11_self_attn_q_proj_weight_to_fp16, x = obj_47_cast_fp16)[name = tensor<string, []>("query_45_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> query_45_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_45_dilations_0, groups = query_45_groups_0, pad = query_45_pad_0, pad_type = query_45_pad_type_0, strides = query_45_strides_0, weight = layers_11_self_attn_q_proj_weight_to_fp16, x = obj_47_cast_fp16)[name = tensor<string, []>("query_45_cast_fp16")];
             tensor<string, []> key_23_pad_type_0 = const()[name = tensor<string, []>("key_23_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> key_23_strides_0 = const()[name = tensor<string, []>("key_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> key_23_pad_0 = const()[name = tensor<string, []>("key_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
@@ -2092,173 +2094,173 @@ program(1.0)
             tensor<int32, [2]> value_23_dilations_0 = const()[name = tensor<string, []>("value_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> value_23_groups_0 = const()[name = tensor<string, []>("value_23_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_11_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(584592128)))];
-            tensor<fp16, [1, 1024, 1, 188]> value_23_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = value_23_dilations_0, groups = value_23_groups_0, pad = value_23_pad_0, pad_type = value_23_pad_type_0, strides = value_23_strides_0, weight = layers_11_self_attn_v_proj_weight_to_fp16, x = obj_47_cast_fp16)[name = tensor<string, []>("value_23_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_3342_to_fp16 = const()[name = tensor<string, []>("op_3342_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(586689344)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_47_cast_fp16 = add(x = query_45_cast_fp16, y = var_3342_to_fp16)[name = tensor<string, []>("query_47_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_3345_to_fp16 = const()[name = tensor<string, []>("op_3345_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(586691456)))];
-            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_23_cast_fp16 = add(x = query_45_cast_fp16, y = var_3345_to_fp16)[name = tensor<string, []>("q_with_bias_v_23_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> value_23_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_23_dilations_0, groups = value_23_groups_0, pad = value_23_pad_0, pad_type = value_23_pad_type_0, strides = value_23_strides_0, weight = layers_11_self_attn_v_proj_weight_to_fp16, x = obj_47_cast_fp16)[name = tensor<string, []>("value_23_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_3340_to_fp16 = const()[name = tensor<string, []>("op_3340_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(586689344)))];
+            tensor<fp16, [1, 1024, 1, 188]> query_47_cast_fp16 = add(x = query_45_cast_fp16, y = var_3340_to_fp16)[name = tensor<string, []>("query_47_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_3343_to_fp16 = const()[name = tensor<string, []>("op_3343_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(586691456)))];
+            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_23_cast_fp16 = add(x = query_45_cast_fp16, y = var_3343_to_fp16)[name = tensor<string, []>("q_with_bias_v_23_cast_fp16")];
             tensor<string, []> p_23_pad_type_0 = const()[name = tensor<string, []>("p_23_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> p_23_strides_0 = const()[name = tensor<string, []>("p_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> p_23_pad_0 = const()[name = tensor<string, []>("p_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> p_23_dilations_0 = const()[name = tensor<string, []>("p_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> p_23_groups_0 = const()[name = tensor<string, []>("p_23_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_11_self_attn_linear_pos_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_linear_pos_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(586693568)))];
-            tensor<fp16, [1, 1024, 1, 375]> p_23_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = p_23_dilations_0, groups = p_23_groups_0, pad = p_23_pad_0, pad_type = p_23_pad_type_0, strides = p_23_strides_0, weight = layers_11_self_attn_linear_pos_weight_to_fp16, x = pos_enc_to_fp16)[name = tensor<string, []>("p_23_cast_fp16")];
-            tensor<int32, [4]> var_3356 = const()[name = tensor<string, []>("op_3356"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_3357_cast_fp16 = reshape(shape = var_3356, x = q_with_bias_v_23_cast_fp16)[name = tensor<string, []>("op_3357_cast_fp16")];
-            tensor<int32, [4]> var_3358 = const()[name = tensor<string, []>("op_3358"), val = tensor<int32, [4]>([1, 8, 128, -1])];
-            tensor<fp16, [1, 8, 128, 375]> var_3359_cast_fp16 = reshape(shape = var_3358, x = p_23_cast_fp16)[name = tensor<string, []>("op_3359_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 375]> p_23_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_23_dilations_0, groups = p_23_groups_0, pad = p_23_pad_0, pad_type = p_23_pad_type_0, strides = p_23_strides_0, weight = layers_11_self_attn_linear_pos_weight_to_fp16, x = obj_3_cast_fp16)[name = tensor<string, []>("p_23_cast_fp16")];
+            tensor<int32, [4]> var_3354 = const()[name = tensor<string, []>("op_3354"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_3355_cast_fp16 = reshape(shape = var_3354, x = q_with_bias_v_23_cast_fp16)[name = tensor<string, []>("op_3355_cast_fp16")];
+            tensor<int32, [4]> var_3356 = const()[name = tensor<string, []>("op_3356"), val = tensor<int32, [4]>([1, 8, 128, -1])];
+            tensor<fp16, [1, 8, 128, 375]> var_3357_cast_fp16 = reshape(shape = var_3356, x = p_23_cast_fp16)[name = tensor<string, []>("op_3357_cast_fp16")];
             tensor<bool, []> matrix_bd_89_transpose_x_0 = const()[name = tensor<string, []>("matrix_bd_89_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> matrix_bd_89_transpose_y_0 = const()[name = tensor<string, []>("matrix_bd_89_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_89_cast_fp16 = matmul(transpose_x = matrix_bd_89_transpose_x_0, transpose_y = matrix_bd_89_transpose_y_0, x = var_3357_cast_fp16, y = var_3359_cast_fp16)[name = tensor<string, []>("matrix_bd_89_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_89_cast_fp16 = matmul(transpose_x = matrix_bd_89_transpose_x_0, transpose_y = matrix_bd_89_transpose_y_0, x = var_3355_cast_fp16, y = var_3357_cast_fp16)[name = tensor<string, []>("matrix_bd_89_cast_fp16")];
             tensor<int32, [8]> matrix_bd_91_pad_0 = const()[name = tensor<string, []>("matrix_bd_91_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
             tensor<string, []> matrix_bd_91_mode_0 = const()[name = tensor<string, []>("matrix_bd_91_mode_0"), val = tensor<string, []>("constant")];
             tensor<fp16, []> const_131_to_fp16 = const()[name = tensor<string, []>("const_131_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
             tensor<fp16, [1, 8, 188, 376]> matrix_bd_91_cast_fp16 = pad(constant_val = const_131_to_fp16, mode = matrix_bd_91_mode_0, pad = matrix_bd_91_pad_0, x = matrix_bd_89_cast_fp16)[name = tensor<string, []>("matrix_bd_91_cast_fp16")];
-            tensor<int32, [4]> var_3368 = const()[name = tensor<string, []>("op_3368"), val = tensor<int32, [4]>([1, 8, -1, 188])];
-            tensor<fp16, [1, 8, 376, 188]> matrix_bd_93_cast_fp16 = reshape(shape = var_3368, x = matrix_bd_91_cast_fp16)[name = tensor<string, []>("matrix_bd_93_cast_fp16")];
-            tensor<int32, [4]> var_3372_begin_0 = const()[name = tensor<string, []>("op_3372_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
-            tensor<int32, [4]> var_3372_end_0 = const()[name = tensor<string, []>("op_3372_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
-            tensor<bool, [4]> var_3372_end_mask_0 = const()[name = tensor<string, []>("op_3372_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
-            tensor<fp16, [1, 8, 375, 188]> var_3372_cast_fp16 = slice_by_index(begin = var_3372_begin_0, end = var_3372_end_0, end_mask = var_3372_end_mask_0, x = matrix_bd_93_cast_fp16)[name = tensor<string, []>("op_3372_cast_fp16")];
-            tensor<int32, [4]> var_3373 = const()[name = tensor<string, []>("op_3373"), val = tensor<int32, [4]>([1, 8, 188, 375])];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_95_cast_fp16 = reshape(shape = var_3373, x = var_3372_cast_fp16)[name = tensor<string, []>("matrix_bd_95_cast_fp16")];
-            tensor<int32, [4]> var_3378_begin_0 = const()[name = tensor<string, []>("op_3378_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [4]> var_3378_end_0 = const()[name = tensor<string, []>("op_3378_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
-            tensor<bool, [4]> var_3378_end_mask_0 = const()[name = tensor<string, []>("op_3378_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
-            tensor<fp16, [1, 8, 188, 188]> var_3378_cast_fp16 = slice_by_index(begin = var_3378_begin_0, end = var_3378_end_0, end_mask = var_3378_end_mask_0, x = matrix_bd_95_cast_fp16)[name = tensor<string, []>("op_3378_cast_fp16")];
-            tensor<fp16, []> var_3379_to_fp16 = const()[name = tensor<string, []>("op_3379_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 188, 188]> qk_mask_23_cast_fp16 = mul(x = var_3378_cast_fp16, y = var_3379_to_fp16)[name = tensor<string, []>("qk_mask_23_cast_fp16")];
-            tensor<int32, [4]> var_3383 = const()[name = tensor<string, []>("op_3383"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> mh_q_23_cast_fp16 = reshape(shape = var_3383, x = query_47_cast_fp16)[name = tensor<string, []>("mh_q_23_cast_fp16")];
-            tensor<fp16, []> var_3385_to_fp16 = const()[name = tensor<string, []>("op_3385_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 128, 188]> var_3386_cast_fp16 = mul(x = mh_q_23_cast_fp16, y = var_3385_to_fp16)[name = tensor<string, []>("op_3386_cast_fp16")];
-            tensor<int32, [4]> var_3389 = const()[name = tensor<string, []>("op_3389"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_3390_cast_fp16 = reshape(shape = var_3389, x = key_23_cast_fp16)[name = tensor<string, []>("op_3390_cast_fp16")];
+            tensor<int32, [4]> var_3366 = const()[name = tensor<string, []>("op_3366"), val = tensor<int32, [4]>([1, 8, -1, 188])];
+            tensor<fp16, [1, 8, 376, 188]> matrix_bd_93_cast_fp16 = reshape(shape = var_3366, x = matrix_bd_91_cast_fp16)[name = tensor<string, []>("matrix_bd_93_cast_fp16")];
+            tensor<int32, [4]> var_3370_begin_0 = const()[name = tensor<string, []>("op_3370_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3370_end_0 = const()[name = tensor<string, []>("op_3370_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
+            tensor<bool, [4]> var_3370_end_mask_0 = const()[name = tensor<string, []>("op_3370_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 375, 188]> var_3370_cast_fp16 = slice_by_index(begin = var_3370_begin_0, end = var_3370_end_0, end_mask = var_3370_end_mask_0, x = matrix_bd_93_cast_fp16)[name = tensor<string, []>("op_3370_cast_fp16")];
+            tensor<int32, [4]> var_3371 = const()[name = tensor<string, []>("op_3371"), val = tensor<int32, [4]>([1, 8, 188, 375])];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_95_cast_fp16 = reshape(shape = var_3371, x = var_3370_cast_fp16)[name = tensor<string, []>("matrix_bd_95_cast_fp16")];
+            tensor<int32, [4]> var_3376_begin_0 = const()[name = tensor<string, []>("op_3376_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_3376_end_0 = const()[name = tensor<string, []>("op_3376_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
+            tensor<bool, [4]> var_3376_end_mask_0 = const()[name = tensor<string, []>("op_3376_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 188, 188]> var_3376_cast_fp16 = slice_by_index(begin = var_3376_begin_0, end = var_3376_end_0, end_mask = var_3376_end_mask_0, x = matrix_bd_95_cast_fp16)[name = tensor<string, []>("op_3376_cast_fp16")];
+            tensor<fp16, []> var_3377_to_fp16 = const()[name = tensor<string, []>("op_3377_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 188, 188]> qk_mask_23_cast_fp16 = mul(x = var_3376_cast_fp16, y = var_3377_to_fp16)[name = tensor<string, []>("qk_mask_23_cast_fp16")];
+            tensor<int32, [4]> var_3381 = const()[name = tensor<string, []>("op_3381"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> mh_q_23_cast_fp16 = reshape(shape = var_3381, x = query_47_cast_fp16)[name = tensor<string, []>("mh_q_23_cast_fp16")];
+            tensor<fp16, []> var_3383_to_fp16 = const()[name = tensor<string, []>("op_3383_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 128, 188]> var_3384_cast_fp16 = mul(x = mh_q_23_cast_fp16, y = var_3383_to_fp16)[name = tensor<string, []>("op_3384_cast_fp16")];
+            tensor<int32, [4]> var_3387 = const()[name = tensor<string, []>("op_3387"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_3388_cast_fp16 = reshape(shape = var_3387, x = key_23_cast_fp16)[name = tensor<string, []>("op_3388_cast_fp16")];
             tensor<bool, []> mh_w_45_transpose_x_0 = const()[name = tensor<string, []>("mh_w_45_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> mh_w_45_transpose_y_0 = const()[name = tensor<string, []>("mh_w_45_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 188]> mh_w_45_cast_fp16 = matmul(transpose_x = mh_w_45_transpose_x_0, transpose_y = mh_w_45_transpose_y_0, x = var_3386_cast_fp16, y = var_3390_cast_fp16)[name = tensor<string, []>("mh_w_45_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> mh_w_45_cast_fp16 = matmul(transpose_x = mh_w_45_transpose_x_0, transpose_y = mh_w_45_transpose_y_0, x = var_3384_cast_fp16, y = var_3388_cast_fp16)[name = tensor<string, []>("mh_w_45_cast_fp16")];
             tensor<fp16, [1, 8, 188, 188]> mh_w_47_cast_fp16 = add(x = mh_w_45_cast_fp16, y = qk_mask_23_cast_fp16)[name = tensor<string, []>("mh_w_47_cast_fp16")];
-            tensor<fp16, [1, 8, 188, 188]> var_3394_cast_fp16 = softmax(axis = var_3235, x = mh_w_47_cast_fp16)[name = tensor<string, []>("op_3394_cast_fp16")];
-            tensor<int32, [4]> var_3395 = const()[name = tensor<string, []>("op_3395"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_3396_cast_fp16 = reshape(shape = var_3395, x = value_23_cast_fp16)[name = tensor<string, []>("op_3396_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> var_3392_cast_fp16 = softmax(axis = var_3233, x = mh_w_47_cast_fp16)[name = tensor<string, []>("op_3392_cast_fp16")];
+            tensor<int32, [4]> var_3393 = const()[name = tensor<string, []>("op_3393"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_3394_cast_fp16 = reshape(shape = var_3393, x = value_23_cast_fp16)[name = tensor<string, []>("op_3394_cast_fp16")];
             tensor<bool, []> attn_23_transpose_x_0 = const()[name = tensor<string, []>("attn_23_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> attn_23_transpose_y_0 = const()[name = tensor<string, []>("attn_23_transpose_y_0"), val = tensor<bool, []>(true)];
-            tensor<fp16, [1, 8, 128, 188]> attn_23_cast_fp16 = matmul(transpose_x = attn_23_transpose_x_0, transpose_y = attn_23_transpose_y_0, x = var_3396_cast_fp16, y = var_3394_cast_fp16)[name = tensor<string, []>("attn_23_cast_fp16")];
-            tensor<int32, [4]> var_3399 = const()[name = tensor<string, []>("op_3399"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
-            tensor<fp16, [1, 1024, 1, 188]> input_307_cast_fp16 = reshape(shape = var_3399, x = attn_23_cast_fp16)[name = tensor<string, []>("input_307_cast_fp16")];
+            tensor<fp16, [1, 8, 128, 188]> attn_23_cast_fp16 = matmul(transpose_x = attn_23_transpose_x_0, transpose_y = attn_23_transpose_y_0, x = var_3394_cast_fp16, y = var_3392_cast_fp16)[name = tensor<string, []>("attn_23_cast_fp16")];
+            tensor<int32, [4]> var_3397 = const()[name = tensor<string, []>("op_3397"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
+            tensor<fp16, [1, 1024, 1, 188]> input_309_cast_fp16 = reshape(shape = var_3397, x = attn_23_cast_fp16)[name = tensor<string, []>("input_309_cast_fp16")];
             tensor<string, []> obj_49_pad_type_0 = const()[name = tensor<string, []>("obj_49_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> obj_49_strides_0 = const()[name = tensor<string, []>("obj_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> obj_49_pad_0 = const()[name = tensor<string, []>("obj_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> obj_49_dilations_0 = const()[name = tensor<string, []>("obj_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> obj_49_groups_0 = const()[name = tensor<string, []>("obj_49_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_11_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(588790784)))];
-            tensor<fp16, [1, 1024, 1, 188]> obj_49_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = obj_49_dilations_0, groups = obj_49_groups_0, pad = obj_49_pad_0, pad_type = obj_49_pad_type_0, strides = obj_49_strides_0, weight = layers_11_self_attn_o_proj_weight_to_fp16, x = input_307_cast_fp16)[name = tensor<string, []>("obj_49_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_49_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_49_dilations_0, groups = obj_49_groups_0, pad = obj_49_pad_0, pad_type = obj_49_pad_type_0, strides = obj_49_strides_0, weight = layers_11_self_attn_o_proj_weight_to_fp16, x = input_309_cast_fp16)[name = tensor<string, []>("obj_49_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_115_cast_fp16 = add(x = inputs_113_cast_fp16, y = obj_49_cast_fp16)[name = tensor<string, []>("inputs_115_cast_fp16")];
             tensor<int32, [1]> out_115_axes_0 = const()[name = tensor<string, []>("out_115_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_3417_to_fp16 = const()[name = tensor<string, []>("op_3417_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_115_cast_fp16 = layer_norm(axes = out_115_axes_0, epsilon = var_3417_to_fp16, x = inputs_115_cast_fp16)[name = tensor<string, []>("out_115_cast_fp16")];
-            tensor<fp16, [1024]> input_309_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_309_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(590888000)))];
-            tensor<fp16, [1024]> input_309_beta_0_to_fp16 = const()[name = tensor<string, []>("input_309_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(590890112)))];
-            tensor<fp16, []> input_309_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_309_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_309_cast_fp16 = batch_norm(beta = input_309_beta_0_to_fp16, epsilon = input_309_epsilon_0_to_fp16, gamma = input_309_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_115_cast_fp16)[name = tensor<string, []>("input_309_cast_fp16")];
-            tensor<string, []> input_311_pad_type_0 = const()[name = tensor<string, []>("input_311_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_311_strides_0 = const()[name = tensor<string, []>("input_311_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_311_pad_0 = const()[name = tensor<string, []>("input_311_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_311_dilations_0 = const()[name = tensor<string, []>("input_311_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_311_groups_0 = const()[name = tensor<string, []>("input_311_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_3415_to_fp16 = const()[name = tensor<string, []>("op_3415_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_115_cast_fp16 = layer_norm(axes = out_115_axes_0, epsilon = var_3415_to_fp16, x = inputs_115_cast_fp16)[name = tensor<string, []>("out_115_cast_fp16")];
+            tensor<fp16, [1024]> input_311_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_311_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(590888000)))];
+            tensor<fp16, [1024]> input_311_beta_0_to_fp16 = const()[name = tensor<string, []>("input_311_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(590890112)))];
+            tensor<fp16, []> input_311_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_311_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_311_cast_fp16 = batch_norm(beta = input_311_beta_0_to_fp16, epsilon = input_311_epsilon_0_to_fp16, gamma = input_311_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_115_cast_fp16)[name = tensor<string, []>("input_311_cast_fp16")];
+            tensor<string, []> input_313_pad_type_0 = const()[name = tensor<string, []>("input_313_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_313_strides_0 = const()[name = tensor<string, []>("input_313_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_313_pad_0 = const()[name = tensor<string, []>("input_313_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_313_dilations_0 = const()[name = tensor<string, []>("input_313_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_313_groups_0 = const()[name = tensor<string, []>("input_313_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [2048, 1024, 1, 1]> layers_11_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(590892224)))];
-            tensor<fp16, [1, 2048, 1, 188]> input_311_cast_fp16 = conv(dilations = input_311_dilations_0, groups = input_311_groups_0, pad = input_311_pad_0, pad_type = input_311_pad_type_0, strides = input_311_strides_0, weight = layers_11_conv_pointwise_conv1_weight_to_fp16, x = input_309_cast_fp16)[name = tensor<string, []>("input_311_cast_fp16")];
-            tensor<int32, []> input_313_split_num_splits_0 = const()[name = tensor<string, []>("input_313_split_num_splits_0"), val = tensor<int32, []>(2)];
-            tensor<int32, []> input_313_split_axis_0 = const()[name = tensor<string, []>("input_313_split_axis_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1, 1024, 1, 188]> input_313_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_313_split_cast_fp16_1 = split(axis = input_313_split_axis_0, num_splits = input_313_split_num_splits_0, x = input_311_cast_fp16)[name = tensor<string, []>("input_313_split_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_313_split_1_sigmoid_cast_fp16 = sigmoid(x = input_313_split_cast_fp16_1)[name = tensor<string, []>("input_313_split_1_sigmoid_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_313_cast_fp16 = mul(x = input_313_split_cast_fp16_0, y = input_313_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_313_cast_fp16")];
-            tensor<string, []> input_315_pad_type_0 = const()[name = tensor<string, []>("input_315_pad_type_0"), val = tensor<string, []>("custom")];
-            tensor<int32, [4]> input_315_pad_0 = const()[name = tensor<string, []>("input_315_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
-            tensor<int32, []> input_315_groups_0 = const()[name = tensor<string, []>("input_315_groups_0"), val = tensor<int32, []>(1024)];
-            tensor<int32, [2]> input_315_strides_0 = const()[name = tensor<string, []>("input_315_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> input_315_dilations_0 = const()[name = tensor<string, []>("input_315_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [1, 2048, 1, 188]> input_313_cast_fp16 = conv(dilations = input_313_dilations_0, groups = input_313_groups_0, pad = input_313_pad_0, pad_type = input_313_pad_type_0, strides = input_313_strides_0, weight = layers_11_conv_pointwise_conv1_weight_to_fp16, x = input_311_cast_fp16)[name = tensor<string, []>("input_313_cast_fp16")];
+            tensor<int32, []> input_315_split_num_splits_0 = const()[name = tensor<string, []>("input_315_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> input_315_split_axis_0 = const()[name = tensor<string, []>("input_315_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 1, 188]> input_315_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_315_split_cast_fp16_1 = split(axis = input_315_split_axis_0, num_splits = input_315_split_num_splits_0, x = input_313_cast_fp16)[name = tensor<string, []>("input_315_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_315_split_1_sigmoid_cast_fp16 = sigmoid(x = input_315_split_cast_fp16_1)[name = tensor<string, []>("input_315_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_315_cast_fp16 = mul(x = input_315_split_cast_fp16_0, y = input_315_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_315_cast_fp16")];
+            tensor<string, []> input_317_pad_type_0 = const()[name = tensor<string, []>("input_317_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_317_pad_0 = const()[name = tensor<string, []>("input_317_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
+            tensor<int32, []> input_317_groups_0 = const()[name = tensor<string, []>("input_317_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [2]> input_317_strides_0 = const()[name = tensor<string, []>("input_317_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> input_317_dilations_0 = const()[name = tensor<string, []>("input_317_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<fp16, [1024, 1, 1, 9]> const_290_to_fp16 = const()[name = tensor<string, []>("const_290_to_fp16"), val = tensor<fp16, [1024, 1, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(595086592)))];
             tensor<fp16, [1024]> const_291_to_fp16 = const()[name = tensor<string, []>("const_291_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(595105088)))];
-            tensor<fp16, [1, 1024, 1, 188]> input_317_cast_fp16 = conv(bias = const_291_to_fp16, dilations = input_315_dilations_0, groups = input_315_groups_0, pad = input_315_pad_0, pad_type = input_315_pad_type_0, strides = input_315_strides_0, weight = const_290_to_fp16, x = input_313_cast_fp16)[name = tensor<string, []>("input_317_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_319_cast_fp16 = silu(x = input_317_cast_fp16)[name = tensor<string, []>("input_319_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_319_cast_fp16 = conv(bias = const_291_to_fp16, dilations = input_317_dilations_0, groups = input_317_groups_0, pad = input_317_pad_0, pad_type = input_317_pad_type_0, strides = input_317_strides_0, weight = const_290_to_fp16, x = input_315_cast_fp16)[name = tensor<string, []>("input_319_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_321_cast_fp16 = silu(x = input_319_cast_fp16)[name = tensor<string, []>("input_321_cast_fp16")];
             tensor<string, []> x_71_pad_type_0 = const()[name = tensor<string, []>("x_71_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_71_strides_0 = const()[name = tensor<string, []>("x_71_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_71_pad_0 = const()[name = tensor<string, []>("x_71_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_71_dilations_0 = const()[name = tensor<string, []>("x_71_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_71_groups_0 = const()[name = tensor<string, []>("x_71_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_11_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(595107200)))];
-            tensor<fp16, [1, 1024, 1, 188]> x_71_cast_fp16 = conv(dilations = x_71_dilations_0, groups = x_71_groups_0, pad = x_71_pad_0, pad_type = x_71_pad_type_0, strides = x_71_strides_0, weight = layers_11_conv_pointwise_conv2_weight_to_fp16, x = input_319_cast_fp16)[name = tensor<string, []>("x_71_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> x_71_cast_fp16 = conv(dilations = x_71_dilations_0, groups = x_71_groups_0, pad = x_71_pad_0, pad_type = x_71_pad_type_0, strides = x_71_strides_0, weight = layers_11_conv_pointwise_conv2_weight_to_fp16, x = input_321_cast_fp16)[name = tensor<string, []>("x_71_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_117_cast_fp16 = add(x = inputs_115_cast_fp16, y = x_71_cast_fp16)[name = tensor<string, []>("inputs_117_cast_fp16")];
             tensor<int32, [1]> out_117_axes_0 = const()[name = tensor<string, []>("out_117_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_3465_to_fp16 = const()[name = tensor<string, []>("op_3465_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_117_cast_fp16 = layer_norm(axes = out_117_axes_0, epsilon = var_3465_to_fp16, x = inputs_117_cast_fp16)[name = tensor<string, []>("out_117_cast_fp16")];
-            tensor<fp16, [1024]> input_321_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_321_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(597204416)))];
-            tensor<fp16, [1024]> input_321_beta_0_to_fp16 = const()[name = tensor<string, []>("input_321_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(597206528)))];
-            tensor<fp16, []> input_321_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_321_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_321_cast_fp16 = batch_norm(beta = input_321_beta_0_to_fp16, epsilon = input_321_epsilon_0_to_fp16, gamma = input_321_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_117_cast_fp16)[name = tensor<string, []>("input_321_cast_fp16")];
-            tensor<string, []> input_323_pad_type_0 = const()[name = tensor<string, []>("input_323_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_323_strides_0 = const()[name = tensor<string, []>("input_323_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_323_pad_0 = const()[name = tensor<string, []>("input_323_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_323_dilations_0 = const()[name = tensor<string, []>("input_323_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_323_groups_0 = const()[name = tensor<string, []>("input_323_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_3463_to_fp16 = const()[name = tensor<string, []>("op_3463_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_117_cast_fp16 = layer_norm(axes = out_117_axes_0, epsilon = var_3463_to_fp16, x = inputs_117_cast_fp16)[name = tensor<string, []>("out_117_cast_fp16")];
+            tensor<fp16, [1024]> input_323_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_323_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(597204416)))];
+            tensor<fp16, [1024]> input_323_beta_0_to_fp16 = const()[name = tensor<string, []>("input_323_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(597206528)))];
+            tensor<fp16, []> input_323_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_323_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_323_cast_fp16 = batch_norm(beta = input_323_beta_0_to_fp16, epsilon = input_323_epsilon_0_to_fp16, gamma = input_323_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_117_cast_fp16)[name = tensor<string, []>("input_323_cast_fp16")];
+            tensor<string, []> input_325_pad_type_0 = const()[name = tensor<string, []>("input_325_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_325_strides_0 = const()[name = tensor<string, []>("input_325_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_325_pad_0 = const()[name = tensor<string, []>("input_325_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_325_dilations_0 = const()[name = tensor<string, []>("input_325_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_325_groups_0 = const()[name = tensor<string, []>("input_325_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_11_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_feed_forward2_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(597208640)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_323_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_323_dilations_0, groups = input_323_groups_0, pad = input_323_pad_0, pad_type = input_323_pad_type_0, strides = input_323_strides_0, weight = layers_11_feed_forward2_fc1_weight_to_fp16, x = input_321_cast_fp16)[name = tensor<string, []>("input_323_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_325_cast_fp16 = silu(x = input_323_cast_fp16)[name = tensor<string, []>("input_325_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_325_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_325_dilations_0, groups = input_325_groups_0, pad = input_325_pad_0, pad_type = input_325_pad_type_0, strides = input_325_strides_0, weight = layers_11_feed_forward2_fc1_weight_to_fp16, x = input_323_cast_fp16)[name = tensor<string, []>("input_325_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_327_cast_fp16 = silu(x = input_325_cast_fp16)[name = tensor<string, []>("input_327_cast_fp16")];
             tensor<string, []> x_73_pad_type_0 = const()[name = tensor<string, []>("x_73_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_73_strides_0 = const()[name = tensor<string, []>("x_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_73_pad_0 = const()[name = tensor<string, []>("x_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_73_dilations_0 = const()[name = tensor<string, []>("x_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_73_groups_0 = const()[name = tensor<string, []>("x_73_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_3493_weight_0_to_fp16 = const()[name = tensor<string, []>("op_3493_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(605597312)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_3493_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_73_dilations_0, groups = x_73_groups_0, pad = x_73_pad_0, pad_type = x_73_pad_type_0, strides = x_73_strides_0, weight = var_3493_weight_0_to_fp16, x = input_325_cast_fp16)[name = tensor<string, []>("op_3493_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_119_cast_fp16 = add(x = inputs_117_cast_fp16, y = var_3493_cast_fp16)[name = tensor<string, []>("inputs_119_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_3491_weight_0_to_fp16 = const()[name = tensor<string, []>("op_3491_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(605597312)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_3491_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_73_dilations_0, groups = x_73_groups_0, pad = x_73_pad_0, pad_type = x_73_pad_type_0, strides = x_73_strides_0, weight = var_3491_weight_0_to_fp16, x = input_327_cast_fp16)[name = tensor<string, []>("op_3491_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_119_cast_fp16 = add(x = inputs_117_cast_fp16, y = var_3491_cast_fp16)[name = tensor<string, []>("inputs_119_cast_fp16")];
             tensor<int32, [1]> out_119_axes_0 = const()[name = tensor<string, []>("out_119_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_3503_to_fp16 = const()[name = tensor<string, []>("op_3503_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_119_cast_fp16 = layer_norm(axes = out_119_axes_0, epsilon = var_3503_to_fp16, x = inputs_119_cast_fp16)[name = tensor<string, []>("out_119_cast_fp16")];
+            tensor<fp16, []> var_3501_to_fp16 = const()[name = tensor<string, []>("op_3501_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_119_cast_fp16 = layer_norm(axes = out_119_axes_0, epsilon = var_3501_to_fp16, x = inputs_119_cast_fp16)[name = tensor<string, []>("out_119_cast_fp16")];
             tensor<fp16, [1024]> inputs_121_gamma_0_to_fp16 = const()[name = tensor<string, []>("inputs_121_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(613985984)))];
             tensor<fp16, [1024]> inputs_121_beta_0_to_fp16 = const()[name = tensor<string, []>("inputs_121_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(613988096)))];
             tensor<fp16, []> inputs_121_epsilon_0_to_fp16 = const()[name = tensor<string, []>("inputs_121_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_121_cast_fp16 = batch_norm(beta = inputs_121_beta_0_to_fp16, epsilon = inputs_121_epsilon_0_to_fp16, gamma = inputs_121_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_119_cast_fp16)[name = tensor<string, []>("inputs_121_cast_fp16")];
-            tensor<int32, []> var_3517 = const()[name = tensor<string, []>("op_3517"), val = tensor<int32, []>(3)];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_121_cast_fp16 = batch_norm(beta = inputs_121_beta_0_to_fp16, epsilon = inputs_121_epsilon_0_to_fp16, gamma = inputs_121_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_119_cast_fp16)[name = tensor<string, []>("inputs_121_cast_fp16")];
+            tensor<int32, []> var_3515 = const()[name = tensor<string, []>("op_3515"), val = tensor<int32, []>(3)];
             tensor<int32, [1]> out_121_axes_0 = const()[name = tensor<string, []>("out_121_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_3548_to_fp16 = const()[name = tensor<string, []>("op_3548_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_121_cast_fp16 = layer_norm(axes = out_121_axes_0, epsilon = var_3548_to_fp16, x = inputs_121_cast_fp16)[name = tensor<string, []>("out_121_cast_fp16")];
-            tensor<fp16, [1024]> input_327_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_327_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(613990208)))];
-            tensor<fp16, [1024]> input_327_beta_0_to_fp16 = const()[name = tensor<string, []>("input_327_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(613992320)))];
-            tensor<fp16, []> input_327_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_327_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_327_cast_fp16 = batch_norm(beta = input_327_beta_0_to_fp16, epsilon = input_327_epsilon_0_to_fp16, gamma = input_327_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_121_cast_fp16)[name = tensor<string, []>("input_327_cast_fp16")];
-            tensor<string, []> input_329_pad_type_0 = const()[name = tensor<string, []>("input_329_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_329_strides_0 = const()[name = tensor<string, []>("input_329_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_329_pad_0 = const()[name = tensor<string, []>("input_329_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_329_dilations_0 = const()[name = tensor<string, []>("input_329_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_329_groups_0 = const()[name = tensor<string, []>("input_329_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_3546_to_fp16 = const()[name = tensor<string, []>("op_3546_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_121_cast_fp16 = layer_norm(axes = out_121_axes_0, epsilon = var_3546_to_fp16, x = inputs_121_cast_fp16)[name = tensor<string, []>("out_121_cast_fp16")];
+            tensor<fp16, [1024]> input_329_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_329_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(613990208)))];
+            tensor<fp16, [1024]> input_329_beta_0_to_fp16 = const()[name = tensor<string, []>("input_329_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(613992320)))];
+            tensor<fp16, []> input_329_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_329_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_329_cast_fp16 = batch_norm(beta = input_329_beta_0_to_fp16, epsilon = input_329_epsilon_0_to_fp16, gamma = input_329_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_121_cast_fp16)[name = tensor<string, []>("input_329_cast_fp16")];
+            tensor<string, []> input_331_pad_type_0 = const()[name = tensor<string, []>("input_331_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_331_strides_0 = const()[name = tensor<string, []>("input_331_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_331_pad_0 = const()[name = tensor<string, []>("input_331_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_331_dilations_0 = const()[name = tensor<string, []>("input_331_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_331_groups_0 = const()[name = tensor<string, []>("input_331_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_12_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_12_feed_forward1_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(613994432)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_329_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_329_dilations_0, groups = input_329_groups_0, pad = input_329_pad_0, pad_type = input_329_pad_type_0, strides = input_329_strides_0, weight = layers_12_feed_forward1_fc1_weight_to_fp16, x = input_327_cast_fp16)[name = tensor<string, []>("input_329_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_331_cast_fp16 = silu(x = input_329_cast_fp16)[name = tensor<string, []>("input_331_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_331_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_331_dilations_0, groups = input_331_groups_0, pad = input_331_pad_0, pad_type = input_331_pad_type_0, strides = input_331_strides_0, weight = layers_12_feed_forward1_fc1_weight_to_fp16, x = input_329_cast_fp16)[name = tensor<string, []>("input_331_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_333_cast_fp16 = silu(x = input_331_cast_fp16)[name = tensor<string, []>("input_333_cast_fp16")];
             tensor<string, []> x_75_pad_type_0 = const()[name = tensor<string, []>("x_75_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_75_strides_0 = const()[name = tensor<string, []>("x_75_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_75_pad_0 = const()[name = tensor<string, []>("x_75_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_75_dilations_0 = const()[name = tensor<string, []>("x_75_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_75_groups_0 = const()[name = tensor<string, []>("x_75_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_3576_weight_0_to_fp16 = const()[name = tensor<string, []>("op_3576_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(622383104)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_3576_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_75_dilations_0, groups = x_75_groups_0, pad = x_75_pad_0, pad_type = x_75_pad_type_0, strides = x_75_strides_0, weight = var_3576_weight_0_to_fp16, x = input_331_cast_fp16)[name = tensor<string, []>("op_3576_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_123_cast_fp16 = add(x = inputs_121_cast_fp16, y = var_3576_cast_fp16)[name = tensor<string, []>("inputs_123_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_3574_weight_0_to_fp16 = const()[name = tensor<string, []>("op_3574_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(622383104)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_3574_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_75_dilations_0, groups = x_75_groups_0, pad = x_75_pad_0, pad_type = x_75_pad_type_0, strides = x_75_strides_0, weight = var_3574_weight_0_to_fp16, x = input_333_cast_fp16)[name = tensor<string, []>("op_3574_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_123_cast_fp16 = add(x = inputs_121_cast_fp16, y = var_3574_cast_fp16)[name = tensor<string, []>("inputs_123_cast_fp16")];
             tensor<int32, [1]> out_123_axes_0 = const()[name = tensor<string, []>("out_123_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_3586_to_fp16 = const()[name = tensor<string, []>("op_3586_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_123_cast_fp16 = layer_norm(axes = out_123_axes_0, epsilon = var_3586_to_fp16, x = inputs_123_cast_fp16)[name = tensor<string, []>("out_123_cast_fp16")];
+            tensor<fp16, []> var_3584_to_fp16 = const()[name = tensor<string, []>("op_3584_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_123_cast_fp16 = layer_norm(axes = out_123_axes_0, epsilon = var_3584_to_fp16, x = inputs_123_cast_fp16)[name = tensor<string, []>("out_123_cast_fp16")];
             tensor<fp16, [1024]> obj_51_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_51_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(630771776)))];
             tensor<fp16, [1024]> obj_51_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_51_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(630773888)))];
             tensor<fp16, []> obj_51_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_51_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> obj_51_cast_fp16 = batch_norm(beta = obj_51_beta_0_to_fp16, epsilon = obj_51_epsilon_0_to_fp16, gamma = obj_51_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_123_cast_fp16)[name = tensor<string, []>("obj_51_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_51_cast_fp16 = batch_norm(beta = obj_51_beta_0_to_fp16, epsilon = obj_51_epsilon_0_to_fp16, gamma = obj_51_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_123_cast_fp16)[name = tensor<string, []>("obj_51_cast_fp16")];
             tensor<string, []> query_49_pad_type_0 = const()[name = tensor<string, []>("query_49_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> query_49_strides_0 = const()[name = tensor<string, []>("query_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> query_49_pad_0 = const()[name = tensor<string, []>("query_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> query_49_dilations_0 = const()[name = tensor<string, []>("query_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> query_49_groups_0 = const()[name = tensor<string, []>("query_49_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_12_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_12_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(630776000)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_49_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = query_49_dilations_0, groups = query_49_groups_0, pad = query_49_pad_0, pad_type = query_49_pad_type_0, strides = query_49_strides_0, weight = layers_12_self_attn_q_proj_weight_to_fp16, x = obj_51_cast_fp16)[name = tensor<string, []>("query_49_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> query_49_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_49_dilations_0, groups = query_49_groups_0, pad = query_49_pad_0, pad_type = query_49_pad_type_0, strides = query_49_strides_0, weight = layers_12_self_attn_q_proj_weight_to_fp16, x = obj_51_cast_fp16)[name = tensor<string, []>("query_49_cast_fp16")];
             tensor<string, []> key_25_pad_type_0 = const()[name = tensor<string, []>("key_25_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> key_25_strides_0 = const()[name = tensor<string, []>("key_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> key_25_pad_0 = const()[name = tensor<string, []>("key_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
@@ -2272,173 +2274,173 @@ program(1.0)
             tensor<int32, [2]> value_25_dilations_0 = const()[name = tensor<string, []>("value_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> value_25_groups_0 = const()[name = tensor<string, []>("value_25_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_12_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_12_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(634970432)))];
-            tensor<fp16, [1, 1024, 1, 188]> value_25_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = value_25_dilations_0, groups = value_25_groups_0, pad = value_25_pad_0, pad_type = value_25_pad_type_0, strides = value_25_strides_0, weight = layers_12_self_attn_v_proj_weight_to_fp16, x = obj_51_cast_fp16)[name = tensor<string, []>("value_25_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_3624_to_fp16 = const()[name = tensor<string, []>("op_3624_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(637067648)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_51_cast_fp16 = add(x = query_49_cast_fp16, y = var_3624_to_fp16)[name = tensor<string, []>("query_51_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_3627_to_fp16 = const()[name = tensor<string, []>("op_3627_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(637069760)))];
-            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_25_cast_fp16 = add(x = query_49_cast_fp16, y = var_3627_to_fp16)[name = tensor<string, []>("q_with_bias_v_25_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> value_25_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_25_dilations_0, groups = value_25_groups_0, pad = value_25_pad_0, pad_type = value_25_pad_type_0, strides = value_25_strides_0, weight = layers_12_self_attn_v_proj_weight_to_fp16, x = obj_51_cast_fp16)[name = tensor<string, []>("value_25_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_3622_to_fp16 = const()[name = tensor<string, []>("op_3622_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(637067648)))];
+            tensor<fp16, [1, 1024, 1, 188]> query_51_cast_fp16 = add(x = query_49_cast_fp16, y = var_3622_to_fp16)[name = tensor<string, []>("query_51_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_3625_to_fp16 = const()[name = tensor<string, []>("op_3625_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(637069760)))];
+            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_25_cast_fp16 = add(x = query_49_cast_fp16, y = var_3625_to_fp16)[name = tensor<string, []>("q_with_bias_v_25_cast_fp16")];
             tensor<string, []> p_25_pad_type_0 = const()[name = tensor<string, []>("p_25_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> p_25_strides_0 = const()[name = tensor<string, []>("p_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> p_25_pad_0 = const()[name = tensor<string, []>("p_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> p_25_dilations_0 = const()[name = tensor<string, []>("p_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> p_25_groups_0 = const()[name = tensor<string, []>("p_25_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_12_self_attn_linear_pos_weight_to_fp16 = const()[name = tensor<string, []>("layers_12_self_attn_linear_pos_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(637071872)))];
-            tensor<fp16, [1, 1024, 1, 375]> p_25_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = p_25_dilations_0, groups = p_25_groups_0, pad = p_25_pad_0, pad_type = p_25_pad_type_0, strides = p_25_strides_0, weight = layers_12_self_attn_linear_pos_weight_to_fp16, x = pos_enc_to_fp16)[name = tensor<string, []>("p_25_cast_fp16")];
-            tensor<int32, [4]> var_3638 = const()[name = tensor<string, []>("op_3638"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_3639_cast_fp16 = reshape(shape = var_3638, x = q_with_bias_v_25_cast_fp16)[name = tensor<string, []>("op_3639_cast_fp16")];
-            tensor<int32, [4]> var_3640 = const()[name = tensor<string, []>("op_3640"), val = tensor<int32, [4]>([1, 8, 128, -1])];
-            tensor<fp16, [1, 8, 128, 375]> var_3641_cast_fp16 = reshape(shape = var_3640, x = p_25_cast_fp16)[name = tensor<string, []>("op_3641_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 375]> p_25_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_25_dilations_0, groups = p_25_groups_0, pad = p_25_pad_0, pad_type = p_25_pad_type_0, strides = p_25_strides_0, weight = layers_12_self_attn_linear_pos_weight_to_fp16, x = obj_3_cast_fp16)[name = tensor<string, []>("p_25_cast_fp16")];
+            tensor<int32, [4]> var_3636 = const()[name = tensor<string, []>("op_3636"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_3637_cast_fp16 = reshape(shape = var_3636, x = q_with_bias_v_25_cast_fp16)[name = tensor<string, []>("op_3637_cast_fp16")];
+            tensor<int32, [4]> var_3638 = const()[name = tensor<string, []>("op_3638"), val = tensor<int32, [4]>([1, 8, 128, -1])];
+            tensor<fp16, [1, 8, 128, 375]> var_3639_cast_fp16 = reshape(shape = var_3638, x = p_25_cast_fp16)[name = tensor<string, []>("op_3639_cast_fp16")];
             tensor<bool, []> matrix_bd_97_transpose_x_0 = const()[name = tensor<string, []>("matrix_bd_97_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> matrix_bd_97_transpose_y_0 = const()[name = tensor<string, []>("matrix_bd_97_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_97_cast_fp16 = matmul(transpose_x = matrix_bd_97_transpose_x_0, transpose_y = matrix_bd_97_transpose_y_0, x = var_3639_cast_fp16, y = var_3641_cast_fp16)[name = tensor<string, []>("matrix_bd_97_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_97_cast_fp16 = matmul(transpose_x = matrix_bd_97_transpose_x_0, transpose_y = matrix_bd_97_transpose_y_0, x = var_3637_cast_fp16, y = var_3639_cast_fp16)[name = tensor<string, []>("matrix_bd_97_cast_fp16")];
             tensor<int32, [8]> matrix_bd_99_pad_0 = const()[name = tensor<string, []>("matrix_bd_99_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
             tensor<string, []> matrix_bd_99_mode_0 = const()[name = tensor<string, []>("matrix_bd_99_mode_0"), val = tensor<string, []>("constant")];
             tensor<fp16, []> const_142_to_fp16 = const()[name = tensor<string, []>("const_142_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
             tensor<fp16, [1, 8, 188, 376]> matrix_bd_99_cast_fp16 = pad(constant_val = const_142_to_fp16, mode = matrix_bd_99_mode_0, pad = matrix_bd_99_pad_0, x = matrix_bd_97_cast_fp16)[name = tensor<string, []>("matrix_bd_99_cast_fp16")];
-            tensor<int32, [4]> var_3650 = const()[name = tensor<string, []>("op_3650"), val = tensor<int32, [4]>([1, 8, -1, 188])];
-            tensor<fp16, [1, 8, 376, 188]> matrix_bd_101_cast_fp16 = reshape(shape = var_3650, x = matrix_bd_99_cast_fp16)[name = tensor<string, []>("matrix_bd_101_cast_fp16")];
-            tensor<int32, [4]> var_3654_begin_0 = const()[name = tensor<string, []>("op_3654_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
-            tensor<int32, [4]> var_3654_end_0 = const()[name = tensor<string, []>("op_3654_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
-            tensor<bool, [4]> var_3654_end_mask_0 = const()[name = tensor<string, []>("op_3654_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
-            tensor<fp16, [1, 8, 375, 188]> var_3654_cast_fp16 = slice_by_index(begin = var_3654_begin_0, end = var_3654_end_0, end_mask = var_3654_end_mask_0, x = matrix_bd_101_cast_fp16)[name = tensor<string, []>("op_3654_cast_fp16")];
-            tensor<int32, [4]> var_3655 = const()[name = tensor<string, []>("op_3655"), val = tensor<int32, [4]>([1, 8, 188, 375])];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_103_cast_fp16 = reshape(shape = var_3655, x = var_3654_cast_fp16)[name = tensor<string, []>("matrix_bd_103_cast_fp16")];
-            tensor<int32, [4]> var_3660_begin_0 = const()[name = tensor<string, []>("op_3660_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [4]> var_3660_end_0 = const()[name = tensor<string, []>("op_3660_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
-            tensor<bool, [4]> var_3660_end_mask_0 = const()[name = tensor<string, []>("op_3660_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
-            tensor<fp16, [1, 8, 188, 188]> var_3660_cast_fp16 = slice_by_index(begin = var_3660_begin_0, end = var_3660_end_0, end_mask = var_3660_end_mask_0, x = matrix_bd_103_cast_fp16)[name = tensor<string, []>("op_3660_cast_fp16")];
-            tensor<fp16, []> var_3661_to_fp16 = const()[name = tensor<string, []>("op_3661_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 188, 188]> qk_mask_25_cast_fp16 = mul(x = var_3660_cast_fp16, y = var_3661_to_fp16)[name = tensor<string, []>("qk_mask_25_cast_fp16")];
-            tensor<int32, [4]> var_3665 = const()[name = tensor<string, []>("op_3665"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> mh_q_25_cast_fp16 = reshape(shape = var_3665, x = query_51_cast_fp16)[name = tensor<string, []>("mh_q_25_cast_fp16")];
-            tensor<fp16, []> var_3667_to_fp16 = const()[name = tensor<string, []>("op_3667_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 128, 188]> var_3668_cast_fp16 = mul(x = mh_q_25_cast_fp16, y = var_3667_to_fp16)[name = tensor<string, []>("op_3668_cast_fp16")];
-            tensor<int32, [4]> var_3671 = const()[name = tensor<string, []>("op_3671"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_3672_cast_fp16 = reshape(shape = var_3671, x = key_25_cast_fp16)[name = tensor<string, []>("op_3672_cast_fp16")];
+            tensor<int32, [4]> var_3648 = const()[name = tensor<string, []>("op_3648"), val = tensor<int32, [4]>([1, 8, -1, 188])];
+            tensor<fp16, [1, 8, 376, 188]> matrix_bd_101_cast_fp16 = reshape(shape = var_3648, x = matrix_bd_99_cast_fp16)[name = tensor<string, []>("matrix_bd_101_cast_fp16")];
+            tensor<int32, [4]> var_3652_begin_0 = const()[name = tensor<string, []>("op_3652_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3652_end_0 = const()[name = tensor<string, []>("op_3652_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
+            tensor<bool, [4]> var_3652_end_mask_0 = const()[name = tensor<string, []>("op_3652_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 375, 188]> var_3652_cast_fp16 = slice_by_index(begin = var_3652_begin_0, end = var_3652_end_0, end_mask = var_3652_end_mask_0, x = matrix_bd_101_cast_fp16)[name = tensor<string, []>("op_3652_cast_fp16")];
+            tensor<int32, [4]> var_3653 = const()[name = tensor<string, []>("op_3653"), val = tensor<int32, [4]>([1, 8, 188, 375])];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_103_cast_fp16 = reshape(shape = var_3653, x = var_3652_cast_fp16)[name = tensor<string, []>("matrix_bd_103_cast_fp16")];
+            tensor<int32, [4]> var_3658_begin_0 = const()[name = tensor<string, []>("op_3658_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_3658_end_0 = const()[name = tensor<string, []>("op_3658_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
+            tensor<bool, [4]> var_3658_end_mask_0 = const()[name = tensor<string, []>("op_3658_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 188, 188]> var_3658_cast_fp16 = slice_by_index(begin = var_3658_begin_0, end = var_3658_end_0, end_mask = var_3658_end_mask_0, x = matrix_bd_103_cast_fp16)[name = tensor<string, []>("op_3658_cast_fp16")];
+            tensor<fp16, []> var_3659_to_fp16 = const()[name = tensor<string, []>("op_3659_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 188, 188]> qk_mask_25_cast_fp16 = mul(x = var_3658_cast_fp16, y = var_3659_to_fp16)[name = tensor<string, []>("qk_mask_25_cast_fp16")];
+            tensor<int32, [4]> var_3663 = const()[name = tensor<string, []>("op_3663"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> mh_q_25_cast_fp16 = reshape(shape = var_3663, x = query_51_cast_fp16)[name = tensor<string, []>("mh_q_25_cast_fp16")];
+            tensor<fp16, []> var_3665_to_fp16 = const()[name = tensor<string, []>("op_3665_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 128, 188]> var_3666_cast_fp16 = mul(x = mh_q_25_cast_fp16, y = var_3665_to_fp16)[name = tensor<string, []>("op_3666_cast_fp16")];
+            tensor<int32, [4]> var_3669 = const()[name = tensor<string, []>("op_3669"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_3670_cast_fp16 = reshape(shape = var_3669, x = key_25_cast_fp16)[name = tensor<string, []>("op_3670_cast_fp16")];
             tensor<bool, []> mh_w_49_transpose_x_0 = const()[name = tensor<string, []>("mh_w_49_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> mh_w_49_transpose_y_0 = const()[name = tensor<string, []>("mh_w_49_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 188]> mh_w_49_cast_fp16 = matmul(transpose_x = mh_w_49_transpose_x_0, transpose_y = mh_w_49_transpose_y_0, x = var_3668_cast_fp16, y = var_3672_cast_fp16)[name = tensor<string, []>("mh_w_49_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> mh_w_49_cast_fp16 = matmul(transpose_x = mh_w_49_transpose_x_0, transpose_y = mh_w_49_transpose_y_0, x = var_3666_cast_fp16, y = var_3670_cast_fp16)[name = tensor<string, []>("mh_w_49_cast_fp16")];
             tensor<fp16, [1, 8, 188, 188]> mh_w_51_cast_fp16 = add(x = mh_w_49_cast_fp16, y = qk_mask_25_cast_fp16)[name = tensor<string, []>("mh_w_51_cast_fp16")];
-            tensor<fp16, [1, 8, 188, 188]> var_3676_cast_fp16 = softmax(axis = var_3517, x = mh_w_51_cast_fp16)[name = tensor<string, []>("op_3676_cast_fp16")];
-            tensor<int32, [4]> var_3677 = const()[name = tensor<string, []>("op_3677"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_3678_cast_fp16 = reshape(shape = var_3677, x = value_25_cast_fp16)[name = tensor<string, []>("op_3678_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> var_3674_cast_fp16 = softmax(axis = var_3515, x = mh_w_51_cast_fp16)[name = tensor<string, []>("op_3674_cast_fp16")];
+            tensor<int32, [4]> var_3675 = const()[name = tensor<string, []>("op_3675"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_3676_cast_fp16 = reshape(shape = var_3675, x = value_25_cast_fp16)[name = tensor<string, []>("op_3676_cast_fp16")];
             tensor<bool, []> attn_25_transpose_x_0 = const()[name = tensor<string, []>("attn_25_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> attn_25_transpose_y_0 = const()[name = tensor<string, []>("attn_25_transpose_y_0"), val = tensor<bool, []>(true)];
-            tensor<fp16, [1, 8, 128, 188]> attn_25_cast_fp16 = matmul(transpose_x = attn_25_transpose_x_0, transpose_y = attn_25_transpose_y_0, x = var_3678_cast_fp16, y = var_3676_cast_fp16)[name = tensor<string, []>("attn_25_cast_fp16")];
-            tensor<int32, [4]> var_3681 = const()[name = tensor<string, []>("op_3681"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
-            tensor<fp16, [1, 1024, 1, 188]> input_333_cast_fp16 = reshape(shape = var_3681, x = attn_25_cast_fp16)[name = tensor<string, []>("input_333_cast_fp16")];
+            tensor<fp16, [1, 8, 128, 188]> attn_25_cast_fp16 = matmul(transpose_x = attn_25_transpose_x_0, transpose_y = attn_25_transpose_y_0, x = var_3676_cast_fp16, y = var_3674_cast_fp16)[name = tensor<string, []>("attn_25_cast_fp16")];
+            tensor<int32, [4]> var_3679 = const()[name = tensor<string, []>("op_3679"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
+            tensor<fp16, [1, 1024, 1, 188]> input_335_cast_fp16 = reshape(shape = var_3679, x = attn_25_cast_fp16)[name = tensor<string, []>("input_335_cast_fp16")];
             tensor<string, []> obj_53_pad_type_0 = const()[name = tensor<string, []>("obj_53_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> obj_53_strides_0 = const()[name = tensor<string, []>("obj_53_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> obj_53_pad_0 = const()[name = tensor<string, []>("obj_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> obj_53_dilations_0 = const()[name = tensor<string, []>("obj_53_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> obj_53_groups_0 = const()[name = tensor<string, []>("obj_53_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_12_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_12_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(639169088)))];
-            tensor<fp16, [1, 1024, 1, 188]> obj_53_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = obj_53_dilations_0, groups = obj_53_groups_0, pad = obj_53_pad_0, pad_type = obj_53_pad_type_0, strides = obj_53_strides_0, weight = layers_12_self_attn_o_proj_weight_to_fp16, x = input_333_cast_fp16)[name = tensor<string, []>("obj_53_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_53_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_53_dilations_0, groups = obj_53_groups_0, pad = obj_53_pad_0, pad_type = obj_53_pad_type_0, strides = obj_53_strides_0, weight = layers_12_self_attn_o_proj_weight_to_fp16, x = input_335_cast_fp16)[name = tensor<string, []>("obj_53_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_125_cast_fp16 = add(x = inputs_123_cast_fp16, y = obj_53_cast_fp16)[name = tensor<string, []>("inputs_125_cast_fp16")];
             tensor<int32, [1]> out_125_axes_0 = const()[name = tensor<string, []>("out_125_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_3699_to_fp16 = const()[name = tensor<string, []>("op_3699_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_125_cast_fp16 = layer_norm(axes = out_125_axes_0, epsilon = var_3699_to_fp16, x = inputs_125_cast_fp16)[name = tensor<string, []>("out_125_cast_fp16")];
-            tensor<fp16, [1024]> input_335_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_335_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(641266304)))];
-            tensor<fp16, [1024]> input_335_beta_0_to_fp16 = const()[name = tensor<string, []>("input_335_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(641268416)))];
-            tensor<fp16, []> input_335_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_335_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_335_cast_fp16 = batch_norm(beta = input_335_beta_0_to_fp16, epsilon = input_335_epsilon_0_to_fp16, gamma = input_335_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_125_cast_fp16)[name = tensor<string, []>("input_335_cast_fp16")];
-            tensor<string, []> input_337_pad_type_0 = const()[name = tensor<string, []>("input_337_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_337_strides_0 = const()[name = tensor<string, []>("input_337_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_337_pad_0 = const()[name = tensor<string, []>("input_337_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_337_dilations_0 = const()[name = tensor<string, []>("input_337_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_337_groups_0 = const()[name = tensor<string, []>("input_337_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_3697_to_fp16 = const()[name = tensor<string, []>("op_3697_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_125_cast_fp16 = layer_norm(axes = out_125_axes_0, epsilon = var_3697_to_fp16, x = inputs_125_cast_fp16)[name = tensor<string, []>("out_125_cast_fp16")];
+            tensor<fp16, [1024]> input_337_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_337_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(641266304)))];
+            tensor<fp16, [1024]> input_337_beta_0_to_fp16 = const()[name = tensor<string, []>("input_337_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(641268416)))];
+            tensor<fp16, []> input_337_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_337_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_337_cast_fp16 = batch_norm(beta = input_337_beta_0_to_fp16, epsilon = input_337_epsilon_0_to_fp16, gamma = input_337_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_125_cast_fp16)[name = tensor<string, []>("input_337_cast_fp16")];
+            tensor<string, []> input_339_pad_type_0 = const()[name = tensor<string, []>("input_339_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_339_strides_0 = const()[name = tensor<string, []>("input_339_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_339_pad_0 = const()[name = tensor<string, []>("input_339_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_339_dilations_0 = const()[name = tensor<string, []>("input_339_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_339_groups_0 = const()[name = tensor<string, []>("input_339_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [2048, 1024, 1, 1]> layers_12_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("layers_12_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(641270528)))];
-            tensor<fp16, [1, 2048, 1, 188]> input_337_cast_fp16 = conv(dilations = input_337_dilations_0, groups = input_337_groups_0, pad = input_337_pad_0, pad_type = input_337_pad_type_0, strides = input_337_strides_0, weight = layers_12_conv_pointwise_conv1_weight_to_fp16, x = input_335_cast_fp16)[name = tensor<string, []>("input_337_cast_fp16")];
-            tensor<int32, []> input_339_split_num_splits_0 = const()[name = tensor<string, []>("input_339_split_num_splits_0"), val = tensor<int32, []>(2)];
-            tensor<int32, []> input_339_split_axis_0 = const()[name = tensor<string, []>("input_339_split_axis_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1, 1024, 1, 188]> input_339_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_339_split_cast_fp16_1 = split(axis = input_339_split_axis_0, num_splits = input_339_split_num_splits_0, x = input_337_cast_fp16)[name = tensor<string, []>("input_339_split_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_339_split_1_sigmoid_cast_fp16 = sigmoid(x = input_339_split_cast_fp16_1)[name = tensor<string, []>("input_339_split_1_sigmoid_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_339_cast_fp16 = mul(x = input_339_split_cast_fp16_0, y = input_339_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_339_cast_fp16")];
-            tensor<string, []> input_341_pad_type_0 = const()[name = tensor<string, []>("input_341_pad_type_0"), val = tensor<string, []>("custom")];
-            tensor<int32, [4]> input_341_pad_0 = const()[name = tensor<string, []>("input_341_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
-            tensor<int32, []> input_341_groups_0 = const()[name = tensor<string, []>("input_341_groups_0"), val = tensor<int32, []>(1024)];
-            tensor<int32, [2]> input_341_strides_0 = const()[name = tensor<string, []>("input_341_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> input_341_dilations_0 = const()[name = tensor<string, []>("input_341_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [1, 2048, 1, 188]> input_339_cast_fp16 = conv(dilations = input_339_dilations_0, groups = input_339_groups_0, pad = input_339_pad_0, pad_type = input_339_pad_type_0, strides = input_339_strides_0, weight = layers_12_conv_pointwise_conv1_weight_to_fp16, x = input_337_cast_fp16)[name = tensor<string, []>("input_339_cast_fp16")];
+            tensor<int32, []> input_341_split_num_splits_0 = const()[name = tensor<string, []>("input_341_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> input_341_split_axis_0 = const()[name = tensor<string, []>("input_341_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 1, 188]> input_341_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_341_split_cast_fp16_1 = split(axis = input_341_split_axis_0, num_splits = input_341_split_num_splits_0, x = input_339_cast_fp16)[name = tensor<string, []>("input_341_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_341_split_1_sigmoid_cast_fp16 = sigmoid(x = input_341_split_cast_fp16_1)[name = tensor<string, []>("input_341_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_341_cast_fp16 = mul(x = input_341_split_cast_fp16_0, y = input_341_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_341_cast_fp16")];
+            tensor<string, []> input_343_pad_type_0 = const()[name = tensor<string, []>("input_343_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_343_pad_0 = const()[name = tensor<string, []>("input_343_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
+            tensor<int32, []> input_343_groups_0 = const()[name = tensor<string, []>("input_343_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [2]> input_343_strides_0 = const()[name = tensor<string, []>("input_343_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> input_343_dilations_0 = const()[name = tensor<string, []>("input_343_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<fp16, [1024, 1, 1, 9]> const_292_to_fp16 = const()[name = tensor<string, []>("const_292_to_fp16"), val = tensor<fp16, [1024, 1, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(645464896)))];
             tensor<fp16, [1024]> const_293_to_fp16 = const()[name = tensor<string, []>("const_293_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(645483392)))];
-            tensor<fp16, [1, 1024, 1, 188]> input_343_cast_fp16 = conv(bias = const_293_to_fp16, dilations = input_341_dilations_0, groups = input_341_groups_0, pad = input_341_pad_0, pad_type = input_341_pad_type_0, strides = input_341_strides_0, weight = const_292_to_fp16, x = input_339_cast_fp16)[name = tensor<string, []>("input_343_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_345_cast_fp16 = silu(x = input_343_cast_fp16)[name = tensor<string, []>("input_345_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_345_cast_fp16 = conv(bias = const_293_to_fp16, dilations = input_343_dilations_0, groups = input_343_groups_0, pad = input_343_pad_0, pad_type = input_343_pad_type_0, strides = input_343_strides_0, weight = const_292_to_fp16, x = input_341_cast_fp16)[name = tensor<string, []>("input_345_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_347_cast_fp16 = silu(x = input_345_cast_fp16)[name = tensor<string, []>("input_347_cast_fp16")];
             tensor<string, []> x_77_pad_type_0 = const()[name = tensor<string, []>("x_77_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_77_strides_0 = const()[name = tensor<string, []>("x_77_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_77_pad_0 = const()[name = tensor<string, []>("x_77_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_77_dilations_0 = const()[name = tensor<string, []>("x_77_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_77_groups_0 = const()[name = tensor<string, []>("x_77_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_12_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("layers_12_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(645485504)))];
-            tensor<fp16, [1, 1024, 1, 188]> x_77_cast_fp16 = conv(dilations = x_77_dilations_0, groups = x_77_groups_0, pad = x_77_pad_0, pad_type = x_77_pad_type_0, strides = x_77_strides_0, weight = layers_12_conv_pointwise_conv2_weight_to_fp16, x = input_345_cast_fp16)[name = tensor<string, []>("x_77_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> x_77_cast_fp16 = conv(dilations = x_77_dilations_0, groups = x_77_groups_0, pad = x_77_pad_0, pad_type = x_77_pad_type_0, strides = x_77_strides_0, weight = layers_12_conv_pointwise_conv2_weight_to_fp16, x = input_347_cast_fp16)[name = tensor<string, []>("x_77_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_127_cast_fp16 = add(x = inputs_125_cast_fp16, y = x_77_cast_fp16)[name = tensor<string, []>("inputs_127_cast_fp16")];
             tensor<int32, [1]> out_127_axes_0 = const()[name = tensor<string, []>("out_127_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_3747_to_fp16 = const()[name = tensor<string, []>("op_3747_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_127_cast_fp16 = layer_norm(axes = out_127_axes_0, epsilon = var_3747_to_fp16, x = inputs_127_cast_fp16)[name = tensor<string, []>("out_127_cast_fp16")];
-            tensor<fp16, [1024]> input_347_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_347_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(647582720)))];
-            tensor<fp16, [1024]> input_347_beta_0_to_fp16 = const()[name = tensor<string, []>("input_347_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(647584832)))];
-            tensor<fp16, []> input_347_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_347_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_347_cast_fp16 = batch_norm(beta = input_347_beta_0_to_fp16, epsilon = input_347_epsilon_0_to_fp16, gamma = input_347_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_127_cast_fp16)[name = tensor<string, []>("input_347_cast_fp16")];
-            tensor<string, []> input_349_pad_type_0 = const()[name = tensor<string, []>("input_349_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_349_strides_0 = const()[name = tensor<string, []>("input_349_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_349_pad_0 = const()[name = tensor<string, []>("input_349_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_349_dilations_0 = const()[name = tensor<string, []>("input_349_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_349_groups_0 = const()[name = tensor<string, []>("input_349_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_3745_to_fp16 = const()[name = tensor<string, []>("op_3745_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_127_cast_fp16 = layer_norm(axes = out_127_axes_0, epsilon = var_3745_to_fp16, x = inputs_127_cast_fp16)[name = tensor<string, []>("out_127_cast_fp16")];
+            tensor<fp16, [1024]> input_349_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_349_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(647582720)))];
+            tensor<fp16, [1024]> input_349_beta_0_to_fp16 = const()[name = tensor<string, []>("input_349_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(647584832)))];
+            tensor<fp16, []> input_349_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_349_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_349_cast_fp16 = batch_norm(beta = input_349_beta_0_to_fp16, epsilon = input_349_epsilon_0_to_fp16, gamma = input_349_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_127_cast_fp16)[name = tensor<string, []>("input_349_cast_fp16")];
+            tensor<string, []> input_351_pad_type_0 = const()[name = tensor<string, []>("input_351_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_351_strides_0 = const()[name = tensor<string, []>("input_351_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_351_pad_0 = const()[name = tensor<string, []>("input_351_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_351_dilations_0 = const()[name = tensor<string, []>("input_351_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_351_groups_0 = const()[name = tensor<string, []>("input_351_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_12_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_12_feed_forward2_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(647586944)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_349_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_349_dilations_0, groups = input_349_groups_0, pad = input_349_pad_0, pad_type = input_349_pad_type_0, strides = input_349_strides_0, weight = layers_12_feed_forward2_fc1_weight_to_fp16, x = input_347_cast_fp16)[name = tensor<string, []>("input_349_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_351_cast_fp16 = silu(x = input_349_cast_fp16)[name = tensor<string, []>("input_351_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_351_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_351_dilations_0, groups = input_351_groups_0, pad = input_351_pad_0, pad_type = input_351_pad_type_0, strides = input_351_strides_0, weight = layers_12_feed_forward2_fc1_weight_to_fp16, x = input_349_cast_fp16)[name = tensor<string, []>("input_351_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_353_cast_fp16 = silu(x = input_351_cast_fp16)[name = tensor<string, []>("input_353_cast_fp16")];
             tensor<string, []> x_79_pad_type_0 = const()[name = tensor<string, []>("x_79_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_79_strides_0 = const()[name = tensor<string, []>("x_79_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_79_pad_0 = const()[name = tensor<string, []>("x_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_79_dilations_0 = const()[name = tensor<string, []>("x_79_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_79_groups_0 = const()[name = tensor<string, []>("x_79_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_3775_weight_0_to_fp16 = const()[name = tensor<string, []>("op_3775_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(655975616)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_3775_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_79_dilations_0, groups = x_79_groups_0, pad = x_79_pad_0, pad_type = x_79_pad_type_0, strides = x_79_strides_0, weight = var_3775_weight_0_to_fp16, x = input_351_cast_fp16)[name = tensor<string, []>("op_3775_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_129_cast_fp16 = add(x = inputs_127_cast_fp16, y = var_3775_cast_fp16)[name = tensor<string, []>("inputs_129_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_3773_weight_0_to_fp16 = const()[name = tensor<string, []>("op_3773_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(655975616)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_3773_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_79_dilations_0, groups = x_79_groups_0, pad = x_79_pad_0, pad_type = x_79_pad_type_0, strides = x_79_strides_0, weight = var_3773_weight_0_to_fp16, x = input_353_cast_fp16)[name = tensor<string, []>("op_3773_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_129_cast_fp16 = add(x = inputs_127_cast_fp16, y = var_3773_cast_fp16)[name = tensor<string, []>("inputs_129_cast_fp16")];
             tensor<int32, [1]> out_129_axes_0 = const()[name = tensor<string, []>("out_129_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_3785_to_fp16 = const()[name = tensor<string, []>("op_3785_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_129_cast_fp16 = layer_norm(axes = out_129_axes_0, epsilon = var_3785_to_fp16, x = inputs_129_cast_fp16)[name = tensor<string, []>("out_129_cast_fp16")];
+            tensor<fp16, []> var_3783_to_fp16 = const()[name = tensor<string, []>("op_3783_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_129_cast_fp16 = layer_norm(axes = out_129_axes_0, epsilon = var_3783_to_fp16, x = inputs_129_cast_fp16)[name = tensor<string, []>("out_129_cast_fp16")];
             tensor<fp16, [1024]> inputs_131_gamma_0_to_fp16 = const()[name = tensor<string, []>("inputs_131_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(664364288)))];
             tensor<fp16, [1024]> inputs_131_beta_0_to_fp16 = const()[name = tensor<string, []>("inputs_131_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(664366400)))];
             tensor<fp16, []> inputs_131_epsilon_0_to_fp16 = const()[name = tensor<string, []>("inputs_131_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_131_cast_fp16 = batch_norm(beta = inputs_131_beta_0_to_fp16, epsilon = inputs_131_epsilon_0_to_fp16, gamma = inputs_131_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_129_cast_fp16)[name = tensor<string, []>("inputs_131_cast_fp16")];
-            tensor<int32, []> var_3799 = const()[name = tensor<string, []>("op_3799"), val = tensor<int32, []>(3)];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_131_cast_fp16 = batch_norm(beta = inputs_131_beta_0_to_fp16, epsilon = inputs_131_epsilon_0_to_fp16, gamma = inputs_131_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_129_cast_fp16)[name = tensor<string, []>("inputs_131_cast_fp16")];
+            tensor<int32, []> var_3797 = const()[name = tensor<string, []>("op_3797"), val = tensor<int32, []>(3)];
             tensor<int32, [1]> out_131_axes_0 = const()[name = tensor<string, []>("out_131_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_3830_to_fp16 = const()[name = tensor<string, []>("op_3830_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_131_cast_fp16 = layer_norm(axes = out_131_axes_0, epsilon = var_3830_to_fp16, x = inputs_131_cast_fp16)[name = tensor<string, []>("out_131_cast_fp16")];
-            tensor<fp16, [1024]> input_353_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_353_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(664368512)))];
-            tensor<fp16, [1024]> input_353_beta_0_to_fp16 = const()[name = tensor<string, []>("input_353_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(664370624)))];
-            tensor<fp16, []> input_353_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_353_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_353_cast_fp16 = batch_norm(beta = input_353_beta_0_to_fp16, epsilon = input_353_epsilon_0_to_fp16, gamma = input_353_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_131_cast_fp16)[name = tensor<string, []>("input_353_cast_fp16")];
-            tensor<string, []> input_355_pad_type_0 = const()[name = tensor<string, []>("input_355_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_355_strides_0 = const()[name = tensor<string, []>("input_355_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_355_pad_0 = const()[name = tensor<string, []>("input_355_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_355_dilations_0 = const()[name = tensor<string, []>("input_355_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_355_groups_0 = const()[name = tensor<string, []>("input_355_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_3828_to_fp16 = const()[name = tensor<string, []>("op_3828_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_131_cast_fp16 = layer_norm(axes = out_131_axes_0, epsilon = var_3828_to_fp16, x = inputs_131_cast_fp16)[name = tensor<string, []>("out_131_cast_fp16")];
+            tensor<fp16, [1024]> input_355_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_355_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(664368512)))];
+            tensor<fp16, [1024]> input_355_beta_0_to_fp16 = const()[name = tensor<string, []>("input_355_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(664370624)))];
+            tensor<fp16, []> input_355_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_355_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_355_cast_fp16 = batch_norm(beta = input_355_beta_0_to_fp16, epsilon = input_355_epsilon_0_to_fp16, gamma = input_355_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_131_cast_fp16)[name = tensor<string, []>("input_355_cast_fp16")];
+            tensor<string, []> input_357_pad_type_0 = const()[name = tensor<string, []>("input_357_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_357_strides_0 = const()[name = tensor<string, []>("input_357_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_357_pad_0 = const()[name = tensor<string, []>("input_357_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_357_dilations_0 = const()[name = tensor<string, []>("input_357_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_357_groups_0 = const()[name = tensor<string, []>("input_357_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_13_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_13_feed_forward1_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(664372736)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_355_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_355_dilations_0, groups = input_355_groups_0, pad = input_355_pad_0, pad_type = input_355_pad_type_0, strides = input_355_strides_0, weight = layers_13_feed_forward1_fc1_weight_to_fp16, x = input_353_cast_fp16)[name = tensor<string, []>("input_355_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_357_cast_fp16 = silu(x = input_355_cast_fp16)[name = tensor<string, []>("input_357_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_357_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_357_dilations_0, groups = input_357_groups_0, pad = input_357_pad_0, pad_type = input_357_pad_type_0, strides = input_357_strides_0, weight = layers_13_feed_forward1_fc1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor<string, []>("input_357_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_359_cast_fp16 = silu(x = input_357_cast_fp16)[name = tensor<string, []>("input_359_cast_fp16")];
             tensor<string, []> x_81_pad_type_0 = const()[name = tensor<string, []>("x_81_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_81_strides_0 = const()[name = tensor<string, []>("x_81_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_81_pad_0 = const()[name = tensor<string, []>("x_81_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_81_dilations_0 = const()[name = tensor<string, []>("x_81_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_81_groups_0 = const()[name = tensor<string, []>("x_81_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_3858_weight_0_to_fp16 = const()[name = tensor<string, []>("op_3858_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(672761408)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_3858_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_81_dilations_0, groups = x_81_groups_0, pad = x_81_pad_0, pad_type = x_81_pad_type_0, strides = x_81_strides_0, weight = var_3858_weight_0_to_fp16, x = input_357_cast_fp16)[name = tensor<string, []>("op_3858_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_133_cast_fp16 = add(x = inputs_131_cast_fp16, y = var_3858_cast_fp16)[name = tensor<string, []>("inputs_133_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_3856_weight_0_to_fp16 = const()[name = tensor<string, []>("op_3856_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(672761408)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_3856_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_81_dilations_0, groups = x_81_groups_0, pad = x_81_pad_0, pad_type = x_81_pad_type_0, strides = x_81_strides_0, weight = var_3856_weight_0_to_fp16, x = input_359_cast_fp16)[name = tensor<string, []>("op_3856_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_133_cast_fp16 = add(x = inputs_131_cast_fp16, y = var_3856_cast_fp16)[name = tensor<string, []>("inputs_133_cast_fp16")];
             tensor<int32, [1]> out_133_axes_0 = const()[name = tensor<string, []>("out_133_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_3868_to_fp16 = const()[name = tensor<string, []>("op_3868_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_133_cast_fp16 = layer_norm(axes = out_133_axes_0, epsilon = var_3868_to_fp16, x = inputs_133_cast_fp16)[name = tensor<string, []>("out_133_cast_fp16")];
+            tensor<fp16, []> var_3866_to_fp16 = const()[name = tensor<string, []>("op_3866_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_133_cast_fp16 = layer_norm(axes = out_133_axes_0, epsilon = var_3866_to_fp16, x = inputs_133_cast_fp16)[name = tensor<string, []>("out_133_cast_fp16")];
             tensor<fp16, [1024]> obj_55_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_55_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(681150080)))];
             tensor<fp16, [1024]> obj_55_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_55_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(681152192)))];
             tensor<fp16, []> obj_55_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_55_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> obj_55_cast_fp16 = batch_norm(beta = obj_55_beta_0_to_fp16, epsilon = obj_55_epsilon_0_to_fp16, gamma = obj_55_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_133_cast_fp16)[name = tensor<string, []>("obj_55_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_55_cast_fp16 = batch_norm(beta = obj_55_beta_0_to_fp16, epsilon = obj_55_epsilon_0_to_fp16, gamma = obj_55_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_133_cast_fp16)[name = tensor<string, []>("obj_55_cast_fp16")];
             tensor<string, []> query_53_pad_type_0 = const()[name = tensor<string, []>("query_53_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> query_53_strides_0 = const()[name = tensor<string, []>("query_53_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> query_53_pad_0 = const()[name = tensor<string, []>("query_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> query_53_dilations_0 = const()[name = tensor<string, []>("query_53_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> query_53_groups_0 = const()[name = tensor<string, []>("query_53_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_13_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_13_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(681154304)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_53_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = query_53_dilations_0, groups = query_53_groups_0, pad = query_53_pad_0, pad_type = query_53_pad_type_0, strides = query_53_strides_0, weight = layers_13_self_attn_q_proj_weight_to_fp16, x = obj_55_cast_fp16)[name = tensor<string, []>("query_53_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> query_53_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_53_dilations_0, groups = query_53_groups_0, pad = query_53_pad_0, pad_type = query_53_pad_type_0, strides = query_53_strides_0, weight = layers_13_self_attn_q_proj_weight_to_fp16, x = obj_55_cast_fp16)[name = tensor<string, []>("query_53_cast_fp16")];
             tensor<string, []> key_27_pad_type_0 = const()[name = tensor<string, []>("key_27_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> key_27_strides_0 = const()[name = tensor<string, []>("key_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> key_27_pad_0 = const()[name = tensor<string, []>("key_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
@@ -2452,173 +2454,173 @@ program(1.0)
             tensor<int32, [2]> value_27_dilations_0 = const()[name = tensor<string, []>("value_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> value_27_groups_0 = const()[name = tensor<string, []>("value_27_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_13_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_13_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(685348736)))];
-            tensor<fp16, [1, 1024, 1, 188]> value_27_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = value_27_dilations_0, groups = value_27_groups_0, pad = value_27_pad_0, pad_type = value_27_pad_type_0, strides = value_27_strides_0, weight = layers_13_self_attn_v_proj_weight_to_fp16, x = obj_55_cast_fp16)[name = tensor<string, []>("value_27_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_3906_to_fp16 = const()[name = tensor<string, []>("op_3906_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(687445952)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_55_cast_fp16 = add(x = query_53_cast_fp16, y = var_3906_to_fp16)[name = tensor<string, []>("query_55_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_3909_to_fp16 = const()[name = tensor<string, []>("op_3909_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(687448064)))];
-            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_27_cast_fp16 = add(x = query_53_cast_fp16, y = var_3909_to_fp16)[name = tensor<string, []>("q_with_bias_v_27_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> value_27_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_27_dilations_0, groups = value_27_groups_0, pad = value_27_pad_0, pad_type = value_27_pad_type_0, strides = value_27_strides_0, weight = layers_13_self_attn_v_proj_weight_to_fp16, x = obj_55_cast_fp16)[name = tensor<string, []>("value_27_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_3904_to_fp16 = const()[name = tensor<string, []>("op_3904_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(687445952)))];
+            tensor<fp16, [1, 1024, 1, 188]> query_55_cast_fp16 = add(x = query_53_cast_fp16, y = var_3904_to_fp16)[name = tensor<string, []>("query_55_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_3907_to_fp16 = const()[name = tensor<string, []>("op_3907_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(687448064)))];
+            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_27_cast_fp16 = add(x = query_53_cast_fp16, y = var_3907_to_fp16)[name = tensor<string, []>("q_with_bias_v_27_cast_fp16")];
             tensor<string, []> p_27_pad_type_0 = const()[name = tensor<string, []>("p_27_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> p_27_strides_0 = const()[name = tensor<string, []>("p_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> p_27_pad_0 = const()[name = tensor<string, []>("p_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> p_27_dilations_0 = const()[name = tensor<string, []>("p_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> p_27_groups_0 = const()[name = tensor<string, []>("p_27_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_13_self_attn_linear_pos_weight_to_fp16 = const()[name = tensor<string, []>("layers_13_self_attn_linear_pos_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(687450176)))];
-            tensor<fp16, [1, 1024, 1, 375]> p_27_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = p_27_dilations_0, groups = p_27_groups_0, pad = p_27_pad_0, pad_type = p_27_pad_type_0, strides = p_27_strides_0, weight = layers_13_self_attn_linear_pos_weight_to_fp16, x = pos_enc_to_fp16)[name = tensor<string, []>("p_27_cast_fp16")];
-            tensor<int32, [4]> var_3920 = const()[name = tensor<string, []>("op_3920"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_3921_cast_fp16 = reshape(shape = var_3920, x = q_with_bias_v_27_cast_fp16)[name = tensor<string, []>("op_3921_cast_fp16")];
-            tensor<int32, [4]> var_3922 = const()[name = tensor<string, []>("op_3922"), val = tensor<int32, [4]>([1, 8, 128, -1])];
-            tensor<fp16, [1, 8, 128, 375]> var_3923_cast_fp16 = reshape(shape = var_3922, x = p_27_cast_fp16)[name = tensor<string, []>("op_3923_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 375]> p_27_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_27_dilations_0, groups = p_27_groups_0, pad = p_27_pad_0, pad_type = p_27_pad_type_0, strides = p_27_strides_0, weight = layers_13_self_attn_linear_pos_weight_to_fp16, x = obj_3_cast_fp16)[name = tensor<string, []>("p_27_cast_fp16")];
+            tensor<int32, [4]> var_3918 = const()[name = tensor<string, []>("op_3918"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_3919_cast_fp16 = reshape(shape = var_3918, x = q_with_bias_v_27_cast_fp16)[name = tensor<string, []>("op_3919_cast_fp16")];
+            tensor<int32, [4]> var_3920 = const()[name = tensor<string, []>("op_3920"), val = tensor<int32, [4]>([1, 8, 128, -1])];
+            tensor<fp16, [1, 8, 128, 375]> var_3921_cast_fp16 = reshape(shape = var_3920, x = p_27_cast_fp16)[name = tensor<string, []>("op_3921_cast_fp16")];
             tensor<bool, []> matrix_bd_105_transpose_x_0 = const()[name = tensor<string, []>("matrix_bd_105_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> matrix_bd_105_transpose_y_0 = const()[name = tensor<string, []>("matrix_bd_105_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_105_cast_fp16 = matmul(transpose_x = matrix_bd_105_transpose_x_0, transpose_y = matrix_bd_105_transpose_y_0, x = var_3921_cast_fp16, y = var_3923_cast_fp16)[name = tensor<string, []>("matrix_bd_105_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_105_cast_fp16 = matmul(transpose_x = matrix_bd_105_transpose_x_0, transpose_y = matrix_bd_105_transpose_y_0, x = var_3919_cast_fp16, y = var_3921_cast_fp16)[name = tensor<string, []>("matrix_bd_105_cast_fp16")];
             tensor<int32, [8]> matrix_bd_107_pad_0 = const()[name = tensor<string, []>("matrix_bd_107_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
             tensor<string, []> matrix_bd_107_mode_0 = const()[name = tensor<string, []>("matrix_bd_107_mode_0"), val = tensor<string, []>("constant")];
             tensor<fp16, []> const_153_to_fp16 = const()[name = tensor<string, []>("const_153_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
             tensor<fp16, [1, 8, 188, 376]> matrix_bd_107_cast_fp16 = pad(constant_val = const_153_to_fp16, mode = matrix_bd_107_mode_0, pad = matrix_bd_107_pad_0, x = matrix_bd_105_cast_fp16)[name = tensor<string, []>("matrix_bd_107_cast_fp16")];
-            tensor<int32, [4]> var_3932 = const()[name = tensor<string, []>("op_3932"), val = tensor<int32, [4]>([1, 8, -1, 188])];
-            tensor<fp16, [1, 8, 376, 188]> matrix_bd_109_cast_fp16 = reshape(shape = var_3932, x = matrix_bd_107_cast_fp16)[name = tensor<string, []>("matrix_bd_109_cast_fp16")];
-            tensor<int32, [4]> var_3936_begin_0 = const()[name = tensor<string, []>("op_3936_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
-            tensor<int32, [4]> var_3936_end_0 = const()[name = tensor<string, []>("op_3936_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
-            tensor<bool, [4]> var_3936_end_mask_0 = const()[name = tensor<string, []>("op_3936_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
-            tensor<fp16, [1, 8, 375, 188]> var_3936_cast_fp16 = slice_by_index(begin = var_3936_begin_0, end = var_3936_end_0, end_mask = var_3936_end_mask_0, x = matrix_bd_109_cast_fp16)[name = tensor<string, []>("op_3936_cast_fp16")];
-            tensor<int32, [4]> var_3937 = const()[name = tensor<string, []>("op_3937"), val = tensor<int32, [4]>([1, 8, 188, 375])];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_111_cast_fp16 = reshape(shape = var_3937, x = var_3936_cast_fp16)[name = tensor<string, []>("matrix_bd_111_cast_fp16")];
-            tensor<int32, [4]> var_3942_begin_0 = const()[name = tensor<string, []>("op_3942_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [4]> var_3942_end_0 = const()[name = tensor<string, []>("op_3942_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
-            tensor<bool, [4]> var_3942_end_mask_0 = const()[name = tensor<string, []>("op_3942_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
-            tensor<fp16, [1, 8, 188, 188]> var_3942_cast_fp16 = slice_by_index(begin = var_3942_begin_0, end = var_3942_end_0, end_mask = var_3942_end_mask_0, x = matrix_bd_111_cast_fp16)[name = tensor<string, []>("op_3942_cast_fp16")];
-            tensor<fp16, []> var_3943_to_fp16 = const()[name = tensor<string, []>("op_3943_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 188, 188]> qk_mask_27_cast_fp16 = mul(x = var_3942_cast_fp16, y = var_3943_to_fp16)[name = tensor<string, []>("qk_mask_27_cast_fp16")];
-            tensor<int32, [4]> var_3947 = const()[name = tensor<string, []>("op_3947"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> mh_q_27_cast_fp16 = reshape(shape = var_3947, x = query_55_cast_fp16)[name = tensor<string, []>("mh_q_27_cast_fp16")];
-            tensor<fp16, []> var_3949_to_fp16 = const()[name = tensor<string, []>("op_3949_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 128, 188]> var_3950_cast_fp16 = mul(x = mh_q_27_cast_fp16, y = var_3949_to_fp16)[name = tensor<string, []>("op_3950_cast_fp16")];
-            tensor<int32, [4]> var_3953 = const()[name = tensor<string, []>("op_3953"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_3954_cast_fp16 = reshape(shape = var_3953, x = key_27_cast_fp16)[name = tensor<string, []>("op_3954_cast_fp16")];
+            tensor<int32, [4]> var_3930 = const()[name = tensor<string, []>("op_3930"), val = tensor<int32, [4]>([1, 8, -1, 188])];
+            tensor<fp16, [1, 8, 376, 188]> matrix_bd_109_cast_fp16 = reshape(shape = var_3930, x = matrix_bd_107_cast_fp16)[name = tensor<string, []>("matrix_bd_109_cast_fp16")];
+            tensor<int32, [4]> var_3934_begin_0 = const()[name = tensor<string, []>("op_3934_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3934_end_0 = const()[name = tensor<string, []>("op_3934_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
+            tensor<bool, [4]> var_3934_end_mask_0 = const()[name = tensor<string, []>("op_3934_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 375, 188]> var_3934_cast_fp16 = slice_by_index(begin = var_3934_begin_0, end = var_3934_end_0, end_mask = var_3934_end_mask_0, x = matrix_bd_109_cast_fp16)[name = tensor<string, []>("op_3934_cast_fp16")];
+            tensor<int32, [4]> var_3935 = const()[name = tensor<string, []>("op_3935"), val = tensor<int32, [4]>([1, 8, 188, 375])];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_111_cast_fp16 = reshape(shape = var_3935, x = var_3934_cast_fp16)[name = tensor<string, []>("matrix_bd_111_cast_fp16")];
+            tensor<int32, [4]> var_3940_begin_0 = const()[name = tensor<string, []>("op_3940_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_3940_end_0 = const()[name = tensor<string, []>("op_3940_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
+            tensor<bool, [4]> var_3940_end_mask_0 = const()[name = tensor<string, []>("op_3940_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 188, 188]> var_3940_cast_fp16 = slice_by_index(begin = var_3940_begin_0, end = var_3940_end_0, end_mask = var_3940_end_mask_0, x = matrix_bd_111_cast_fp16)[name = tensor<string, []>("op_3940_cast_fp16")];
+            tensor<fp16, []> var_3941_to_fp16 = const()[name = tensor<string, []>("op_3941_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 188, 188]> qk_mask_27_cast_fp16 = mul(x = var_3940_cast_fp16, y = var_3941_to_fp16)[name = tensor<string, []>("qk_mask_27_cast_fp16")];
+            tensor<int32, [4]> var_3945 = const()[name = tensor<string, []>("op_3945"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> mh_q_27_cast_fp16 = reshape(shape = var_3945, x = query_55_cast_fp16)[name = tensor<string, []>("mh_q_27_cast_fp16")];
+            tensor<fp16, []> var_3947_to_fp16 = const()[name = tensor<string, []>("op_3947_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 128, 188]> var_3948_cast_fp16 = mul(x = mh_q_27_cast_fp16, y = var_3947_to_fp16)[name = tensor<string, []>("op_3948_cast_fp16")];
+            tensor<int32, [4]> var_3951 = const()[name = tensor<string, []>("op_3951"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_3952_cast_fp16 = reshape(shape = var_3951, x = key_27_cast_fp16)[name = tensor<string, []>("op_3952_cast_fp16")];
             tensor<bool, []> mh_w_53_transpose_x_0 = const()[name = tensor<string, []>("mh_w_53_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> mh_w_53_transpose_y_0 = const()[name = tensor<string, []>("mh_w_53_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 188]> mh_w_53_cast_fp16 = matmul(transpose_x = mh_w_53_transpose_x_0, transpose_y = mh_w_53_transpose_y_0, x = var_3950_cast_fp16, y = var_3954_cast_fp16)[name = tensor<string, []>("mh_w_53_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> mh_w_53_cast_fp16 = matmul(transpose_x = mh_w_53_transpose_x_0, transpose_y = mh_w_53_transpose_y_0, x = var_3948_cast_fp16, y = var_3952_cast_fp16)[name = tensor<string, []>("mh_w_53_cast_fp16")];
             tensor<fp16, [1, 8, 188, 188]> mh_w_55_cast_fp16 = add(x = mh_w_53_cast_fp16, y = qk_mask_27_cast_fp16)[name = tensor<string, []>("mh_w_55_cast_fp16")];
-            tensor<fp16, [1, 8, 188, 188]> var_3958_cast_fp16 = softmax(axis = var_3799, x = mh_w_55_cast_fp16)[name = tensor<string, []>("op_3958_cast_fp16")];
-            tensor<int32, [4]> var_3959 = const()[name = tensor<string, []>("op_3959"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_3960_cast_fp16 = reshape(shape = var_3959, x = value_27_cast_fp16)[name = tensor<string, []>("op_3960_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> var_3956_cast_fp16 = softmax(axis = var_3797, x = mh_w_55_cast_fp16)[name = tensor<string, []>("op_3956_cast_fp16")];
+            tensor<int32, [4]> var_3957 = const()[name = tensor<string, []>("op_3957"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_3958_cast_fp16 = reshape(shape = var_3957, x = value_27_cast_fp16)[name = tensor<string, []>("op_3958_cast_fp16")];
             tensor<bool, []> attn_27_transpose_x_0 = const()[name = tensor<string, []>("attn_27_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> attn_27_transpose_y_0 = const()[name = tensor<string, []>("attn_27_transpose_y_0"), val = tensor<bool, []>(true)];
-            tensor<fp16, [1, 8, 128, 188]> attn_27_cast_fp16 = matmul(transpose_x = attn_27_transpose_x_0, transpose_y = attn_27_transpose_y_0, x = var_3960_cast_fp16, y = var_3958_cast_fp16)[name = tensor<string, []>("attn_27_cast_fp16")];
-            tensor<int32, [4]> var_3963 = const()[name = tensor<string, []>("op_3963"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
-            tensor<fp16, [1, 1024, 1, 188]> input_359_cast_fp16 = reshape(shape = var_3963, x = attn_27_cast_fp16)[name = tensor<string, []>("input_359_cast_fp16")];
+            tensor<fp16, [1, 8, 128, 188]> attn_27_cast_fp16 = matmul(transpose_x = attn_27_transpose_x_0, transpose_y = attn_27_transpose_y_0, x = var_3958_cast_fp16, y = var_3956_cast_fp16)[name = tensor<string, []>("attn_27_cast_fp16")];
+            tensor<int32, [4]> var_3961 = const()[name = tensor<string, []>("op_3961"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
+            tensor<fp16, [1, 1024, 1, 188]> input_361_cast_fp16 = reshape(shape = var_3961, x = attn_27_cast_fp16)[name = tensor<string, []>("input_361_cast_fp16")];
             tensor<string, []> obj_57_pad_type_0 = const()[name = tensor<string, []>("obj_57_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> obj_57_strides_0 = const()[name = tensor<string, []>("obj_57_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> obj_57_pad_0 = const()[name = tensor<string, []>("obj_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> obj_57_dilations_0 = const()[name = tensor<string, []>("obj_57_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> obj_57_groups_0 = const()[name = tensor<string, []>("obj_57_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_13_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_13_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(689547392)))];
-            tensor<fp16, [1, 1024, 1, 188]> obj_57_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = obj_57_dilations_0, groups = obj_57_groups_0, pad = obj_57_pad_0, pad_type = obj_57_pad_type_0, strides = obj_57_strides_0, weight = layers_13_self_attn_o_proj_weight_to_fp16, x = input_359_cast_fp16)[name = tensor<string, []>("obj_57_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_57_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_57_dilations_0, groups = obj_57_groups_0, pad = obj_57_pad_0, pad_type = obj_57_pad_type_0, strides = obj_57_strides_0, weight = layers_13_self_attn_o_proj_weight_to_fp16, x = input_361_cast_fp16)[name = tensor<string, []>("obj_57_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_135_cast_fp16 = add(x = inputs_133_cast_fp16, y = obj_57_cast_fp16)[name = tensor<string, []>("inputs_135_cast_fp16")];
             tensor<int32, [1]> out_135_axes_0 = const()[name = tensor<string, []>("out_135_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_3981_to_fp16 = const()[name = tensor<string, []>("op_3981_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_135_cast_fp16 = layer_norm(axes = out_135_axes_0, epsilon = var_3981_to_fp16, x = inputs_135_cast_fp16)[name = tensor<string, []>("out_135_cast_fp16")];
-            tensor<fp16, [1024]> input_361_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_361_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(691644608)))];
-            tensor<fp16, [1024]> input_361_beta_0_to_fp16 = const()[name = tensor<string, []>("input_361_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(691646720)))];
-            tensor<fp16, []> input_361_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_361_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_361_cast_fp16 = batch_norm(beta = input_361_beta_0_to_fp16, epsilon = input_361_epsilon_0_to_fp16, gamma = input_361_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_135_cast_fp16)[name = tensor<string, []>("input_361_cast_fp16")];
-            tensor<string, []> input_363_pad_type_0 = const()[name = tensor<string, []>("input_363_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_363_strides_0 = const()[name = tensor<string, []>("input_363_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_363_pad_0 = const()[name = tensor<string, []>("input_363_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_363_dilations_0 = const()[name = tensor<string, []>("input_363_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_363_groups_0 = const()[name = tensor<string, []>("input_363_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_3979_to_fp16 = const()[name = tensor<string, []>("op_3979_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_135_cast_fp16 = layer_norm(axes = out_135_axes_0, epsilon = var_3979_to_fp16, x = inputs_135_cast_fp16)[name = tensor<string, []>("out_135_cast_fp16")];
+            tensor<fp16, [1024]> input_363_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_363_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(691644608)))];
+            tensor<fp16, [1024]> input_363_beta_0_to_fp16 = const()[name = tensor<string, []>("input_363_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(691646720)))];
+            tensor<fp16, []> input_363_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_363_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_363_cast_fp16 = batch_norm(beta = input_363_beta_0_to_fp16, epsilon = input_363_epsilon_0_to_fp16, gamma = input_363_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_135_cast_fp16)[name = tensor<string, []>("input_363_cast_fp16")];
+            tensor<string, []> input_365_pad_type_0 = const()[name = tensor<string, []>("input_365_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_365_strides_0 = const()[name = tensor<string, []>("input_365_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_365_pad_0 = const()[name = tensor<string, []>("input_365_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_365_dilations_0 = const()[name = tensor<string, []>("input_365_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_365_groups_0 = const()[name = tensor<string, []>("input_365_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [2048, 1024, 1, 1]> layers_13_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("layers_13_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(691648832)))];
-            tensor<fp16, [1, 2048, 1, 188]> input_363_cast_fp16 = conv(dilations = input_363_dilations_0, groups = input_363_groups_0, pad = input_363_pad_0, pad_type = input_363_pad_type_0, strides = input_363_strides_0, weight = layers_13_conv_pointwise_conv1_weight_to_fp16, x = input_361_cast_fp16)[name = tensor<string, []>("input_363_cast_fp16")];
-            tensor<int32, []> input_365_split_num_splits_0 = const()[name = tensor<string, []>("input_365_split_num_splits_0"), val = tensor<int32, []>(2)];
-            tensor<int32, []> input_365_split_axis_0 = const()[name = tensor<string, []>("input_365_split_axis_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1, 1024, 1, 188]> input_365_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_365_split_cast_fp16_1 = split(axis = input_365_split_axis_0, num_splits = input_365_split_num_splits_0, x = input_363_cast_fp16)[name = tensor<string, []>("input_365_split_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_365_split_1_sigmoid_cast_fp16 = sigmoid(x = input_365_split_cast_fp16_1)[name = tensor<string, []>("input_365_split_1_sigmoid_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_365_cast_fp16 = mul(x = input_365_split_cast_fp16_0, y = input_365_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_365_cast_fp16")];
-            tensor<string, []> input_367_pad_type_0 = const()[name = tensor<string, []>("input_367_pad_type_0"), val = tensor<string, []>("custom")];
-            tensor<int32, [4]> input_367_pad_0 = const()[name = tensor<string, []>("input_367_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
-            tensor<int32, []> input_367_groups_0 = const()[name = tensor<string, []>("input_367_groups_0"), val = tensor<int32, []>(1024)];
-            tensor<int32, [2]> input_367_strides_0 = const()[name = tensor<string, []>("input_367_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> input_367_dilations_0 = const()[name = tensor<string, []>("input_367_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [1, 2048, 1, 188]> input_365_cast_fp16 = conv(dilations = input_365_dilations_0, groups = input_365_groups_0, pad = input_365_pad_0, pad_type = input_365_pad_type_0, strides = input_365_strides_0, weight = layers_13_conv_pointwise_conv1_weight_to_fp16, x = input_363_cast_fp16)[name = tensor<string, []>("input_365_cast_fp16")];
+            tensor<int32, []> input_367_split_num_splits_0 = const()[name = tensor<string, []>("input_367_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> input_367_split_axis_0 = const()[name = tensor<string, []>("input_367_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 1, 188]> input_367_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_367_split_cast_fp16_1 = split(axis = input_367_split_axis_0, num_splits = input_367_split_num_splits_0, x = input_365_cast_fp16)[name = tensor<string, []>("input_367_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_367_split_1_sigmoid_cast_fp16 = sigmoid(x = input_367_split_cast_fp16_1)[name = tensor<string, []>("input_367_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_367_cast_fp16 = mul(x = input_367_split_cast_fp16_0, y = input_367_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_367_cast_fp16")];
+            tensor<string, []> input_369_pad_type_0 = const()[name = tensor<string, []>("input_369_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_369_pad_0 = const()[name = tensor<string, []>("input_369_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
+            tensor<int32, []> input_369_groups_0 = const()[name = tensor<string, []>("input_369_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [2]> input_369_strides_0 = const()[name = tensor<string, []>("input_369_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> input_369_dilations_0 = const()[name = tensor<string, []>("input_369_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<fp16, [1024, 1, 1, 9]> const_294_to_fp16 = const()[name = tensor<string, []>("const_294_to_fp16"), val = tensor<fp16, [1024, 1, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(695843200)))];
             tensor<fp16, [1024]> const_295_to_fp16 = const()[name = tensor<string, []>("const_295_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(695861696)))];
-            tensor<fp16, [1, 1024, 1, 188]> input_369_cast_fp16 = conv(bias = const_295_to_fp16, dilations = input_367_dilations_0, groups = input_367_groups_0, pad = input_367_pad_0, pad_type = input_367_pad_type_0, strides = input_367_strides_0, weight = const_294_to_fp16, x = input_365_cast_fp16)[name = tensor<string, []>("input_369_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_371_cast_fp16 = silu(x = input_369_cast_fp16)[name = tensor<string, []>("input_371_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_371_cast_fp16 = conv(bias = const_295_to_fp16, dilations = input_369_dilations_0, groups = input_369_groups_0, pad = input_369_pad_0, pad_type = input_369_pad_type_0, strides = input_369_strides_0, weight = const_294_to_fp16, x = input_367_cast_fp16)[name = tensor<string, []>("input_371_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_373_cast_fp16 = silu(x = input_371_cast_fp16)[name = tensor<string, []>("input_373_cast_fp16")];
             tensor<string, []> x_83_pad_type_0 = const()[name = tensor<string, []>("x_83_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_83_strides_0 = const()[name = tensor<string, []>("x_83_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_83_pad_0 = const()[name = tensor<string, []>("x_83_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_83_dilations_0 = const()[name = tensor<string, []>("x_83_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_83_groups_0 = const()[name = tensor<string, []>("x_83_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_13_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("layers_13_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(695863808)))];
-            tensor<fp16, [1, 1024, 1, 188]> x_83_cast_fp16 = conv(dilations = x_83_dilations_0, groups = x_83_groups_0, pad = x_83_pad_0, pad_type = x_83_pad_type_0, strides = x_83_strides_0, weight = layers_13_conv_pointwise_conv2_weight_to_fp16, x = input_371_cast_fp16)[name = tensor<string, []>("x_83_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> x_83_cast_fp16 = conv(dilations = x_83_dilations_0, groups = x_83_groups_0, pad = x_83_pad_0, pad_type = x_83_pad_type_0, strides = x_83_strides_0, weight = layers_13_conv_pointwise_conv2_weight_to_fp16, x = input_373_cast_fp16)[name = tensor<string, []>("x_83_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_137_cast_fp16 = add(x = inputs_135_cast_fp16, y = x_83_cast_fp16)[name = tensor<string, []>("inputs_137_cast_fp16")];
             tensor<int32, [1]> out_137_axes_0 = const()[name = tensor<string, []>("out_137_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_4029_to_fp16 = const()[name = tensor<string, []>("op_4029_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_137_cast_fp16 = layer_norm(axes = out_137_axes_0, epsilon = var_4029_to_fp16, x = inputs_137_cast_fp16)[name = tensor<string, []>("out_137_cast_fp16")];
-            tensor<fp16, [1024]> input_373_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_373_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(697961024)))];
-            tensor<fp16, [1024]> input_373_beta_0_to_fp16 = const()[name = tensor<string, []>("input_373_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(697963136)))];
-            tensor<fp16, []> input_373_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_373_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_373_cast_fp16 = batch_norm(beta = input_373_beta_0_to_fp16, epsilon = input_373_epsilon_0_to_fp16, gamma = input_373_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_137_cast_fp16)[name = tensor<string, []>("input_373_cast_fp16")];
-            tensor<string, []> input_375_pad_type_0 = const()[name = tensor<string, []>("input_375_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_375_strides_0 = const()[name = tensor<string, []>("input_375_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_375_pad_0 = const()[name = tensor<string, []>("input_375_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_375_dilations_0 = const()[name = tensor<string, []>("input_375_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_375_groups_0 = const()[name = tensor<string, []>("input_375_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_4027_to_fp16 = const()[name = tensor<string, []>("op_4027_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_137_cast_fp16 = layer_norm(axes = out_137_axes_0, epsilon = var_4027_to_fp16, x = inputs_137_cast_fp16)[name = tensor<string, []>("out_137_cast_fp16")];
+            tensor<fp16, [1024]> input_375_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_375_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(697961024)))];
+            tensor<fp16, [1024]> input_375_beta_0_to_fp16 = const()[name = tensor<string, []>("input_375_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(697963136)))];
+            tensor<fp16, []> input_375_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_375_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_375_cast_fp16 = batch_norm(beta = input_375_beta_0_to_fp16, epsilon = input_375_epsilon_0_to_fp16, gamma = input_375_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_137_cast_fp16)[name = tensor<string, []>("input_375_cast_fp16")];
+            tensor<string, []> input_377_pad_type_0 = const()[name = tensor<string, []>("input_377_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_377_strides_0 = const()[name = tensor<string, []>("input_377_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_377_pad_0 = const()[name = tensor<string, []>("input_377_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_377_dilations_0 = const()[name = tensor<string, []>("input_377_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_377_groups_0 = const()[name = tensor<string, []>("input_377_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_13_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_13_feed_forward2_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(697965248)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_375_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_375_dilations_0, groups = input_375_groups_0, pad = input_375_pad_0, pad_type = input_375_pad_type_0, strides = input_375_strides_0, weight = layers_13_feed_forward2_fc1_weight_to_fp16, x = input_373_cast_fp16)[name = tensor<string, []>("input_375_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_377_cast_fp16 = silu(x = input_375_cast_fp16)[name = tensor<string, []>("input_377_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_377_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_377_dilations_0, groups = input_377_groups_0, pad = input_377_pad_0, pad_type = input_377_pad_type_0, strides = input_377_strides_0, weight = layers_13_feed_forward2_fc1_weight_to_fp16, x = input_375_cast_fp16)[name = tensor<string, []>("input_377_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_379_cast_fp16 = silu(x = input_377_cast_fp16)[name = tensor<string, []>("input_379_cast_fp16")];
             tensor<string, []> x_85_pad_type_0 = const()[name = tensor<string, []>("x_85_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_85_strides_0 = const()[name = tensor<string, []>("x_85_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_85_pad_0 = const()[name = tensor<string, []>("x_85_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_85_dilations_0 = const()[name = tensor<string, []>("x_85_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_85_groups_0 = const()[name = tensor<string, []>("x_85_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_4057_weight_0_to_fp16 = const()[name = tensor<string, []>("op_4057_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(706353920)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_4057_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_85_dilations_0, groups = x_85_groups_0, pad = x_85_pad_0, pad_type = x_85_pad_type_0, strides = x_85_strides_0, weight = var_4057_weight_0_to_fp16, x = input_377_cast_fp16)[name = tensor<string, []>("op_4057_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_139_cast_fp16 = add(x = inputs_137_cast_fp16, y = var_4057_cast_fp16)[name = tensor<string, []>("inputs_139_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_4055_weight_0_to_fp16 = const()[name = tensor<string, []>("op_4055_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(706353920)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_4055_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_85_dilations_0, groups = x_85_groups_0, pad = x_85_pad_0, pad_type = x_85_pad_type_0, strides = x_85_strides_0, weight = var_4055_weight_0_to_fp16, x = input_379_cast_fp16)[name = tensor<string, []>("op_4055_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_139_cast_fp16 = add(x = inputs_137_cast_fp16, y = var_4055_cast_fp16)[name = tensor<string, []>("inputs_139_cast_fp16")];
             tensor<int32, [1]> out_139_axes_0 = const()[name = tensor<string, []>("out_139_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_4067_to_fp16 = const()[name = tensor<string, []>("op_4067_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_139_cast_fp16 = layer_norm(axes = out_139_axes_0, epsilon = var_4067_to_fp16, x = inputs_139_cast_fp16)[name = tensor<string, []>("out_139_cast_fp16")];
+            tensor<fp16, []> var_4065_to_fp16 = const()[name = tensor<string, []>("op_4065_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_139_cast_fp16 = layer_norm(axes = out_139_axes_0, epsilon = var_4065_to_fp16, x = inputs_139_cast_fp16)[name = tensor<string, []>("out_139_cast_fp16")];
             tensor<fp16, [1024]> inputs_141_gamma_0_to_fp16 = const()[name = tensor<string, []>("inputs_141_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(714742592)))];
             tensor<fp16, [1024]> inputs_141_beta_0_to_fp16 = const()[name = tensor<string, []>("inputs_141_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(714744704)))];
             tensor<fp16, []> inputs_141_epsilon_0_to_fp16 = const()[name = tensor<string, []>("inputs_141_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_141_cast_fp16 = batch_norm(beta = inputs_141_beta_0_to_fp16, epsilon = inputs_141_epsilon_0_to_fp16, gamma = inputs_141_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_139_cast_fp16)[name = tensor<string, []>("inputs_141_cast_fp16")];
-            tensor<int32, []> var_4081 = const()[name = tensor<string, []>("op_4081"), val = tensor<int32, []>(3)];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_141_cast_fp16 = batch_norm(beta = inputs_141_beta_0_to_fp16, epsilon = inputs_141_epsilon_0_to_fp16, gamma = inputs_141_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_139_cast_fp16)[name = tensor<string, []>("inputs_141_cast_fp16")];
+            tensor<int32, []> var_4079 = const()[name = tensor<string, []>("op_4079"), val = tensor<int32, []>(3)];
             tensor<int32, [1]> out_141_axes_0 = const()[name = tensor<string, []>("out_141_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_4112_to_fp16 = const()[name = tensor<string, []>("op_4112_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_141_cast_fp16 = layer_norm(axes = out_141_axes_0, epsilon = var_4112_to_fp16, x = inputs_141_cast_fp16)[name = tensor<string, []>("out_141_cast_fp16")];
-            tensor<fp16, [1024]> input_379_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_379_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(714746816)))];
-            tensor<fp16, [1024]> input_379_beta_0_to_fp16 = const()[name = tensor<string, []>("input_379_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(714748928)))];
-            tensor<fp16, []> input_379_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_379_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_379_cast_fp16 = batch_norm(beta = input_379_beta_0_to_fp16, epsilon = input_379_epsilon_0_to_fp16, gamma = input_379_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_141_cast_fp16)[name = tensor<string, []>("input_379_cast_fp16")];
-            tensor<string, []> input_381_pad_type_0 = const()[name = tensor<string, []>("input_381_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_381_strides_0 = const()[name = tensor<string, []>("input_381_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_381_pad_0 = const()[name = tensor<string, []>("input_381_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_381_dilations_0 = const()[name = tensor<string, []>("input_381_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_381_groups_0 = const()[name = tensor<string, []>("input_381_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_4110_to_fp16 = const()[name = tensor<string, []>("op_4110_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_141_cast_fp16 = layer_norm(axes = out_141_axes_0, epsilon = var_4110_to_fp16, x = inputs_141_cast_fp16)[name = tensor<string, []>("out_141_cast_fp16")];
+            tensor<fp16, [1024]> input_381_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_381_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(714746816)))];
+            tensor<fp16, [1024]> input_381_beta_0_to_fp16 = const()[name = tensor<string, []>("input_381_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(714748928)))];
+            tensor<fp16, []> input_381_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_381_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_381_cast_fp16 = batch_norm(beta = input_381_beta_0_to_fp16, epsilon = input_381_epsilon_0_to_fp16, gamma = input_381_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_141_cast_fp16)[name = tensor<string, []>("input_381_cast_fp16")];
+            tensor<string, []> input_383_pad_type_0 = const()[name = tensor<string, []>("input_383_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_383_strides_0 = const()[name = tensor<string, []>("input_383_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_383_pad_0 = const()[name = tensor<string, []>("input_383_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_383_dilations_0 = const()[name = tensor<string, []>("input_383_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_383_groups_0 = const()[name = tensor<string, []>("input_383_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_14_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_14_feed_forward1_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(714751040)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_381_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_381_dilations_0, groups = input_381_groups_0, pad = input_381_pad_0, pad_type = input_381_pad_type_0, strides = input_381_strides_0, weight = layers_14_feed_forward1_fc1_weight_to_fp16, x = input_379_cast_fp16)[name = tensor<string, []>("input_381_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_383_cast_fp16 = silu(x = input_381_cast_fp16)[name = tensor<string, []>("input_383_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_383_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_383_dilations_0, groups = input_383_groups_0, pad = input_383_pad_0, pad_type = input_383_pad_type_0, strides = input_383_strides_0, weight = layers_14_feed_forward1_fc1_weight_to_fp16, x = input_381_cast_fp16)[name = tensor<string, []>("input_383_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_385_cast_fp16 = silu(x = input_383_cast_fp16)[name = tensor<string, []>("input_385_cast_fp16")];
             tensor<string, []> x_87_pad_type_0 = const()[name = tensor<string, []>("x_87_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_87_strides_0 = const()[name = tensor<string, []>("x_87_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_87_pad_0 = const()[name = tensor<string, []>("x_87_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_87_dilations_0 = const()[name = tensor<string, []>("x_87_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_87_groups_0 = const()[name = tensor<string, []>("x_87_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_4140_weight_0_to_fp16 = const()[name = tensor<string, []>("op_4140_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(723139712)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_4140_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_87_dilations_0, groups = x_87_groups_0, pad = x_87_pad_0, pad_type = x_87_pad_type_0, strides = x_87_strides_0, weight = var_4140_weight_0_to_fp16, x = input_383_cast_fp16)[name = tensor<string, []>("op_4140_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_143_cast_fp16 = add(x = inputs_141_cast_fp16, y = var_4140_cast_fp16)[name = tensor<string, []>("inputs_143_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_4138_weight_0_to_fp16 = const()[name = tensor<string, []>("op_4138_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(723139712)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_4138_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_87_dilations_0, groups = x_87_groups_0, pad = x_87_pad_0, pad_type = x_87_pad_type_0, strides = x_87_strides_0, weight = var_4138_weight_0_to_fp16, x = input_385_cast_fp16)[name = tensor<string, []>("op_4138_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_143_cast_fp16 = add(x = inputs_141_cast_fp16, y = var_4138_cast_fp16)[name = tensor<string, []>("inputs_143_cast_fp16")];
             tensor<int32, [1]> out_143_axes_0 = const()[name = tensor<string, []>("out_143_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_4150_to_fp16 = const()[name = tensor<string, []>("op_4150_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_143_cast_fp16 = layer_norm(axes = out_143_axes_0, epsilon = var_4150_to_fp16, x = inputs_143_cast_fp16)[name = tensor<string, []>("out_143_cast_fp16")];
+            tensor<fp16, []> var_4148_to_fp16 = const()[name = tensor<string, []>("op_4148_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_143_cast_fp16 = layer_norm(axes = out_143_axes_0, epsilon = var_4148_to_fp16, x = inputs_143_cast_fp16)[name = tensor<string, []>("out_143_cast_fp16")];
             tensor<fp16, [1024]> obj_59_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_59_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(731528384)))];
             tensor<fp16, [1024]> obj_59_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_59_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(731530496)))];
             tensor<fp16, []> obj_59_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_59_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> obj_59_cast_fp16 = batch_norm(beta = obj_59_beta_0_to_fp16, epsilon = obj_59_epsilon_0_to_fp16, gamma = obj_59_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_143_cast_fp16)[name = tensor<string, []>("obj_59_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_59_cast_fp16 = batch_norm(beta = obj_59_beta_0_to_fp16, epsilon = obj_59_epsilon_0_to_fp16, gamma = obj_59_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_143_cast_fp16)[name = tensor<string, []>("obj_59_cast_fp16")];
             tensor<string, []> query_57_pad_type_0 = const()[name = tensor<string, []>("query_57_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> query_57_strides_0 = const()[name = tensor<string, []>("query_57_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> query_57_pad_0 = const()[name = tensor<string, []>("query_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> query_57_dilations_0 = const()[name = tensor<string, []>("query_57_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> query_57_groups_0 = const()[name = tensor<string, []>("query_57_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_14_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_14_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(731532608)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_57_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = query_57_dilations_0, groups = query_57_groups_0, pad = query_57_pad_0, pad_type = query_57_pad_type_0, strides = query_57_strides_0, weight = layers_14_self_attn_q_proj_weight_to_fp16, x = obj_59_cast_fp16)[name = tensor<string, []>("query_57_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> query_57_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_57_dilations_0, groups = query_57_groups_0, pad = query_57_pad_0, pad_type = query_57_pad_type_0, strides = query_57_strides_0, weight = layers_14_self_attn_q_proj_weight_to_fp16, x = obj_59_cast_fp16)[name = tensor<string, []>("query_57_cast_fp16")];
             tensor<string, []> key_29_pad_type_0 = const()[name = tensor<string, []>("key_29_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> key_29_strides_0 = const()[name = tensor<string, []>("key_29_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> key_29_pad_0 = const()[name = tensor<string, []>("key_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
@@ -2632,173 +2634,173 @@ program(1.0)
             tensor<int32, [2]> value_29_dilations_0 = const()[name = tensor<string, []>("value_29_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> value_29_groups_0 = const()[name = tensor<string, []>("value_29_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_14_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_14_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(735727040)))];
-            tensor<fp16, [1, 1024, 1, 188]> value_29_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = value_29_dilations_0, groups = value_29_groups_0, pad = value_29_pad_0, pad_type = value_29_pad_type_0, strides = value_29_strides_0, weight = layers_14_self_attn_v_proj_weight_to_fp16, x = obj_59_cast_fp16)[name = tensor<string, []>("value_29_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_4188_to_fp16 = const()[name = tensor<string, []>("op_4188_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(737824256)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_59_cast_fp16 = add(x = query_57_cast_fp16, y = var_4188_to_fp16)[name = tensor<string, []>("query_59_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_4191_to_fp16 = const()[name = tensor<string, []>("op_4191_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(737826368)))];
-            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_29_cast_fp16 = add(x = query_57_cast_fp16, y = var_4191_to_fp16)[name = tensor<string, []>("q_with_bias_v_29_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> value_29_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_29_dilations_0, groups = value_29_groups_0, pad = value_29_pad_0, pad_type = value_29_pad_type_0, strides = value_29_strides_0, weight = layers_14_self_attn_v_proj_weight_to_fp16, x = obj_59_cast_fp16)[name = tensor<string, []>("value_29_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_4186_to_fp16 = const()[name = tensor<string, []>("op_4186_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(737824256)))];
+            tensor<fp16, [1, 1024, 1, 188]> query_59_cast_fp16 = add(x = query_57_cast_fp16, y = var_4186_to_fp16)[name = tensor<string, []>("query_59_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_4189_to_fp16 = const()[name = tensor<string, []>("op_4189_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(737826368)))];
+            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_29_cast_fp16 = add(x = query_57_cast_fp16, y = var_4189_to_fp16)[name = tensor<string, []>("q_with_bias_v_29_cast_fp16")];
             tensor<string, []> p_29_pad_type_0 = const()[name = tensor<string, []>("p_29_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> p_29_strides_0 = const()[name = tensor<string, []>("p_29_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> p_29_pad_0 = const()[name = tensor<string, []>("p_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> p_29_dilations_0 = const()[name = tensor<string, []>("p_29_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> p_29_groups_0 = const()[name = tensor<string, []>("p_29_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_14_self_attn_linear_pos_weight_to_fp16 = const()[name = tensor<string, []>("layers_14_self_attn_linear_pos_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(737828480)))];
-            tensor<fp16, [1, 1024, 1, 375]> p_29_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = p_29_dilations_0, groups = p_29_groups_0, pad = p_29_pad_0, pad_type = p_29_pad_type_0, strides = p_29_strides_0, weight = layers_14_self_attn_linear_pos_weight_to_fp16, x = pos_enc_to_fp16)[name = tensor<string, []>("p_29_cast_fp16")];
-            tensor<int32, [4]> var_4202 = const()[name = tensor<string, []>("op_4202"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_4203_cast_fp16 = reshape(shape = var_4202, x = q_with_bias_v_29_cast_fp16)[name = tensor<string, []>("op_4203_cast_fp16")];
-            tensor<int32, [4]> var_4204 = const()[name = tensor<string, []>("op_4204"), val = tensor<int32, [4]>([1, 8, 128, -1])];
-            tensor<fp16, [1, 8, 128, 375]> var_4205_cast_fp16 = reshape(shape = var_4204, x = p_29_cast_fp16)[name = tensor<string, []>("op_4205_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 375]> p_29_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_29_dilations_0, groups = p_29_groups_0, pad = p_29_pad_0, pad_type = p_29_pad_type_0, strides = p_29_strides_0, weight = layers_14_self_attn_linear_pos_weight_to_fp16, x = obj_3_cast_fp16)[name = tensor<string, []>("p_29_cast_fp16")];
+            tensor<int32, [4]> var_4200 = const()[name = tensor<string, []>("op_4200"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_4201_cast_fp16 = reshape(shape = var_4200, x = q_with_bias_v_29_cast_fp16)[name = tensor<string, []>("op_4201_cast_fp16")];
+            tensor<int32, [4]> var_4202 = const()[name = tensor<string, []>("op_4202"), val = tensor<int32, [4]>([1, 8, 128, -1])];
+            tensor<fp16, [1, 8, 128, 375]> var_4203_cast_fp16 = reshape(shape = var_4202, x = p_29_cast_fp16)[name = tensor<string, []>("op_4203_cast_fp16")];
             tensor<bool, []> matrix_bd_113_transpose_x_0 = const()[name = tensor<string, []>("matrix_bd_113_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> matrix_bd_113_transpose_y_0 = const()[name = tensor<string, []>("matrix_bd_113_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_113_cast_fp16 = matmul(transpose_x = matrix_bd_113_transpose_x_0, transpose_y = matrix_bd_113_transpose_y_0, x = var_4203_cast_fp16, y = var_4205_cast_fp16)[name = tensor<string, []>("matrix_bd_113_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_113_cast_fp16 = matmul(transpose_x = matrix_bd_113_transpose_x_0, transpose_y = matrix_bd_113_transpose_y_0, x = var_4201_cast_fp16, y = var_4203_cast_fp16)[name = tensor<string, []>("matrix_bd_113_cast_fp16")];
             tensor<int32, [8]> matrix_bd_115_pad_0 = const()[name = tensor<string, []>("matrix_bd_115_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
             tensor<string, []> matrix_bd_115_mode_0 = const()[name = tensor<string, []>("matrix_bd_115_mode_0"), val = tensor<string, []>("constant")];
             tensor<fp16, []> const_164_to_fp16 = const()[name = tensor<string, []>("const_164_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
             tensor<fp16, [1, 8, 188, 376]> matrix_bd_115_cast_fp16 = pad(constant_val = const_164_to_fp16, mode = matrix_bd_115_mode_0, pad = matrix_bd_115_pad_0, x = matrix_bd_113_cast_fp16)[name = tensor<string, []>("matrix_bd_115_cast_fp16")];
-            tensor<int32, [4]> var_4214 = const()[name = tensor<string, []>("op_4214"), val = tensor<int32, [4]>([1, 8, -1, 188])];
-            tensor<fp16, [1, 8, 376, 188]> matrix_bd_117_cast_fp16 = reshape(shape = var_4214, x = matrix_bd_115_cast_fp16)[name = tensor<string, []>("matrix_bd_117_cast_fp16")];
-            tensor<int32, [4]> var_4218_begin_0 = const()[name = tensor<string, []>("op_4218_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
-            tensor<int32, [4]> var_4218_end_0 = const()[name = tensor<string, []>("op_4218_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
-            tensor<bool, [4]> var_4218_end_mask_0 = const()[name = tensor<string, []>("op_4218_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
-            tensor<fp16, [1, 8, 375, 188]> var_4218_cast_fp16 = slice_by_index(begin = var_4218_begin_0, end = var_4218_end_0, end_mask = var_4218_end_mask_0, x = matrix_bd_117_cast_fp16)[name = tensor<string, []>("op_4218_cast_fp16")];
-            tensor<int32, [4]> var_4219 = const()[name = tensor<string, []>("op_4219"), val = tensor<int32, [4]>([1, 8, 188, 375])];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_119_cast_fp16 = reshape(shape = var_4219, x = var_4218_cast_fp16)[name = tensor<string, []>("matrix_bd_119_cast_fp16")];
-            tensor<int32, [4]> var_4224_begin_0 = const()[name = tensor<string, []>("op_4224_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [4]> var_4224_end_0 = const()[name = tensor<string, []>("op_4224_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
-            tensor<bool, [4]> var_4224_end_mask_0 = const()[name = tensor<string, []>("op_4224_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
-            tensor<fp16, [1, 8, 188, 188]> var_4224_cast_fp16 = slice_by_index(begin = var_4224_begin_0, end = var_4224_end_0, end_mask = var_4224_end_mask_0, x = matrix_bd_119_cast_fp16)[name = tensor<string, []>("op_4224_cast_fp16")];
-            tensor<fp16, []> var_4225_to_fp16 = const()[name = tensor<string, []>("op_4225_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 188, 188]> qk_mask_29_cast_fp16 = mul(x = var_4224_cast_fp16, y = var_4225_to_fp16)[name = tensor<string, []>("qk_mask_29_cast_fp16")];
-            tensor<int32, [4]> var_4229 = const()[name = tensor<string, []>("op_4229"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> mh_q_29_cast_fp16 = reshape(shape = var_4229, x = query_59_cast_fp16)[name = tensor<string, []>("mh_q_29_cast_fp16")];
-            tensor<fp16, []> var_4231_to_fp16 = const()[name = tensor<string, []>("op_4231_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 128, 188]> var_4232_cast_fp16 = mul(x = mh_q_29_cast_fp16, y = var_4231_to_fp16)[name = tensor<string, []>("op_4232_cast_fp16")];
-            tensor<int32, [4]> var_4235 = const()[name = tensor<string, []>("op_4235"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_4236_cast_fp16 = reshape(shape = var_4235, x = key_29_cast_fp16)[name = tensor<string, []>("op_4236_cast_fp16")];
+            tensor<int32, [4]> var_4212 = const()[name = tensor<string, []>("op_4212"), val = tensor<int32, [4]>([1, 8, -1, 188])];
+            tensor<fp16, [1, 8, 376, 188]> matrix_bd_117_cast_fp16 = reshape(shape = var_4212, x = matrix_bd_115_cast_fp16)[name = tensor<string, []>("matrix_bd_117_cast_fp16")];
+            tensor<int32, [4]> var_4216_begin_0 = const()[name = tensor<string, []>("op_4216_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_4216_end_0 = const()[name = tensor<string, []>("op_4216_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
+            tensor<bool, [4]> var_4216_end_mask_0 = const()[name = tensor<string, []>("op_4216_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 375, 188]> var_4216_cast_fp16 = slice_by_index(begin = var_4216_begin_0, end = var_4216_end_0, end_mask = var_4216_end_mask_0, x = matrix_bd_117_cast_fp16)[name = tensor<string, []>("op_4216_cast_fp16")];
+            tensor<int32, [4]> var_4217 = const()[name = tensor<string, []>("op_4217"), val = tensor<int32, [4]>([1, 8, 188, 375])];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_119_cast_fp16 = reshape(shape = var_4217, x = var_4216_cast_fp16)[name = tensor<string, []>("matrix_bd_119_cast_fp16")];
+            tensor<int32, [4]> var_4222_begin_0 = const()[name = tensor<string, []>("op_4222_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_4222_end_0 = const()[name = tensor<string, []>("op_4222_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
+            tensor<bool, [4]> var_4222_end_mask_0 = const()[name = tensor<string, []>("op_4222_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 188, 188]> var_4222_cast_fp16 = slice_by_index(begin = var_4222_begin_0, end = var_4222_end_0, end_mask = var_4222_end_mask_0, x = matrix_bd_119_cast_fp16)[name = tensor<string, []>("op_4222_cast_fp16")];
+            tensor<fp16, []> var_4223_to_fp16 = const()[name = tensor<string, []>("op_4223_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 188, 188]> qk_mask_29_cast_fp16 = mul(x = var_4222_cast_fp16, y = var_4223_to_fp16)[name = tensor<string, []>("qk_mask_29_cast_fp16")];
+            tensor<int32, [4]> var_4227 = const()[name = tensor<string, []>("op_4227"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> mh_q_29_cast_fp16 = reshape(shape = var_4227, x = query_59_cast_fp16)[name = tensor<string, []>("mh_q_29_cast_fp16")];
+            tensor<fp16, []> var_4229_to_fp16 = const()[name = tensor<string, []>("op_4229_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 128, 188]> var_4230_cast_fp16 = mul(x = mh_q_29_cast_fp16, y = var_4229_to_fp16)[name = tensor<string, []>("op_4230_cast_fp16")];
+            tensor<int32, [4]> var_4233 = const()[name = tensor<string, []>("op_4233"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_4234_cast_fp16 = reshape(shape = var_4233, x = key_29_cast_fp16)[name = tensor<string, []>("op_4234_cast_fp16")];
             tensor<bool, []> mh_w_57_transpose_x_0 = const()[name = tensor<string, []>("mh_w_57_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> mh_w_57_transpose_y_0 = const()[name = tensor<string, []>("mh_w_57_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 188]> mh_w_57_cast_fp16 = matmul(transpose_x = mh_w_57_transpose_x_0, transpose_y = mh_w_57_transpose_y_0, x = var_4232_cast_fp16, y = var_4236_cast_fp16)[name = tensor<string, []>("mh_w_57_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> mh_w_57_cast_fp16 = matmul(transpose_x = mh_w_57_transpose_x_0, transpose_y = mh_w_57_transpose_y_0, x = var_4230_cast_fp16, y = var_4234_cast_fp16)[name = tensor<string, []>("mh_w_57_cast_fp16")];
             tensor<fp16, [1, 8, 188, 188]> mh_w_59_cast_fp16 = add(x = mh_w_57_cast_fp16, y = qk_mask_29_cast_fp16)[name = tensor<string, []>("mh_w_59_cast_fp16")];
-            tensor<fp16, [1, 8, 188, 188]> var_4240_cast_fp16 = softmax(axis = var_4081, x = mh_w_59_cast_fp16)[name = tensor<string, []>("op_4240_cast_fp16")];
-            tensor<int32, [4]> var_4241 = const()[name = tensor<string, []>("op_4241"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_4242_cast_fp16 = reshape(shape = var_4241, x = value_29_cast_fp16)[name = tensor<string, []>("op_4242_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> var_4238_cast_fp16 = softmax(axis = var_4079, x = mh_w_59_cast_fp16)[name = tensor<string, []>("op_4238_cast_fp16")];
+            tensor<int32, [4]> var_4239 = const()[name = tensor<string, []>("op_4239"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_4240_cast_fp16 = reshape(shape = var_4239, x = value_29_cast_fp16)[name = tensor<string, []>("op_4240_cast_fp16")];
             tensor<bool, []> attn_29_transpose_x_0 = const()[name = tensor<string, []>("attn_29_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> attn_29_transpose_y_0 = const()[name = tensor<string, []>("attn_29_transpose_y_0"), val = tensor<bool, []>(true)];
-            tensor<fp16, [1, 8, 128, 188]> attn_29_cast_fp16 = matmul(transpose_x = attn_29_transpose_x_0, transpose_y = attn_29_transpose_y_0, x = var_4242_cast_fp16, y = var_4240_cast_fp16)[name = tensor<string, []>("attn_29_cast_fp16")];
-            tensor<int32, [4]> var_4245 = const()[name = tensor<string, []>("op_4245"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
-            tensor<fp16, [1, 1024, 1, 188]> input_385_cast_fp16 = reshape(shape = var_4245, x = attn_29_cast_fp16)[name = tensor<string, []>("input_385_cast_fp16")];
+            tensor<fp16, [1, 8, 128, 188]> attn_29_cast_fp16 = matmul(transpose_x = attn_29_transpose_x_0, transpose_y = attn_29_transpose_y_0, x = var_4240_cast_fp16, y = var_4238_cast_fp16)[name = tensor<string, []>("attn_29_cast_fp16")];
+            tensor<int32, [4]> var_4243 = const()[name = tensor<string, []>("op_4243"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
+            tensor<fp16, [1, 1024, 1, 188]> input_387_cast_fp16 = reshape(shape = var_4243, x = attn_29_cast_fp16)[name = tensor<string, []>("input_387_cast_fp16")];
             tensor<string, []> obj_61_pad_type_0 = const()[name = tensor<string, []>("obj_61_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> obj_61_strides_0 = const()[name = tensor<string, []>("obj_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> obj_61_pad_0 = const()[name = tensor<string, []>("obj_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> obj_61_dilations_0 = const()[name = tensor<string, []>("obj_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> obj_61_groups_0 = const()[name = tensor<string, []>("obj_61_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_14_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_14_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(739925696)))];
-            tensor<fp16, [1, 1024, 1, 188]> obj_61_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = obj_61_dilations_0, groups = obj_61_groups_0, pad = obj_61_pad_0, pad_type = obj_61_pad_type_0, strides = obj_61_strides_0, weight = layers_14_self_attn_o_proj_weight_to_fp16, x = input_385_cast_fp16)[name = tensor<string, []>("obj_61_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_61_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_61_dilations_0, groups = obj_61_groups_0, pad = obj_61_pad_0, pad_type = obj_61_pad_type_0, strides = obj_61_strides_0, weight = layers_14_self_attn_o_proj_weight_to_fp16, x = input_387_cast_fp16)[name = tensor<string, []>("obj_61_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_145_cast_fp16 = add(x = inputs_143_cast_fp16, y = obj_61_cast_fp16)[name = tensor<string, []>("inputs_145_cast_fp16")];
             tensor<int32, [1]> out_145_axes_0 = const()[name = tensor<string, []>("out_145_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_4263_to_fp16 = const()[name = tensor<string, []>("op_4263_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_145_cast_fp16 = layer_norm(axes = out_145_axes_0, epsilon = var_4263_to_fp16, x = inputs_145_cast_fp16)[name = tensor<string, []>("out_145_cast_fp16")];
-            tensor<fp16, [1024]> input_387_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_387_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(742022912)))];
-            tensor<fp16, [1024]> input_387_beta_0_to_fp16 = const()[name = tensor<string, []>("input_387_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(742025024)))];
-            tensor<fp16, []> input_387_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_387_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_387_cast_fp16 = batch_norm(beta = input_387_beta_0_to_fp16, epsilon = input_387_epsilon_0_to_fp16, gamma = input_387_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_145_cast_fp16)[name = tensor<string, []>("input_387_cast_fp16")];
-            tensor<string, []> input_389_pad_type_0 = const()[name = tensor<string, []>("input_389_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_389_strides_0 = const()[name = tensor<string, []>("input_389_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_389_pad_0 = const()[name = tensor<string, []>("input_389_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_389_dilations_0 = const()[name = tensor<string, []>("input_389_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_389_groups_0 = const()[name = tensor<string, []>("input_389_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_4261_to_fp16 = const()[name = tensor<string, []>("op_4261_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_145_cast_fp16 = layer_norm(axes = out_145_axes_0, epsilon = var_4261_to_fp16, x = inputs_145_cast_fp16)[name = tensor<string, []>("out_145_cast_fp16")];
+            tensor<fp16, [1024]> input_389_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_389_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(742022912)))];
+            tensor<fp16, [1024]> input_389_beta_0_to_fp16 = const()[name = tensor<string, []>("input_389_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(742025024)))];
+            tensor<fp16, []> input_389_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_389_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_389_cast_fp16 = batch_norm(beta = input_389_beta_0_to_fp16, epsilon = input_389_epsilon_0_to_fp16, gamma = input_389_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_145_cast_fp16)[name = tensor<string, []>("input_389_cast_fp16")];
+            tensor<string, []> input_391_pad_type_0 = const()[name = tensor<string, []>("input_391_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_391_strides_0 = const()[name = tensor<string, []>("input_391_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_391_pad_0 = const()[name = tensor<string, []>("input_391_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_391_dilations_0 = const()[name = tensor<string, []>("input_391_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_391_groups_0 = const()[name = tensor<string, []>("input_391_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [2048, 1024, 1, 1]> layers_14_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("layers_14_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(742027136)))];
-            tensor<fp16, [1, 2048, 1, 188]> input_389_cast_fp16 = conv(dilations = input_389_dilations_0, groups = input_389_groups_0, pad = input_389_pad_0, pad_type = input_389_pad_type_0, strides = input_389_strides_0, weight = layers_14_conv_pointwise_conv1_weight_to_fp16, x = input_387_cast_fp16)[name = tensor<string, []>("input_389_cast_fp16")];
-            tensor<int32, []> input_391_split_num_splits_0 = const()[name = tensor<string, []>("input_391_split_num_splits_0"), val = tensor<int32, []>(2)];
-            tensor<int32, []> input_391_split_axis_0 = const()[name = tensor<string, []>("input_391_split_axis_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1, 1024, 1, 188]> input_391_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_391_split_cast_fp16_1 = split(axis = input_391_split_axis_0, num_splits = input_391_split_num_splits_0, x = input_389_cast_fp16)[name = tensor<string, []>("input_391_split_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_391_split_1_sigmoid_cast_fp16 = sigmoid(x = input_391_split_cast_fp16_1)[name = tensor<string, []>("input_391_split_1_sigmoid_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_391_cast_fp16 = mul(x = input_391_split_cast_fp16_0, y = input_391_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_391_cast_fp16")];
-            tensor<string, []> input_393_pad_type_0 = const()[name = tensor<string, []>("input_393_pad_type_0"), val = tensor<string, []>("custom")];
-            tensor<int32, [4]> input_393_pad_0 = const()[name = tensor<string, []>("input_393_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
-            tensor<int32, []> input_393_groups_0 = const()[name = tensor<string, []>("input_393_groups_0"), val = tensor<int32, []>(1024)];
-            tensor<int32, [2]> input_393_strides_0 = const()[name = tensor<string, []>("input_393_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> input_393_dilations_0 = const()[name = tensor<string, []>("input_393_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [1, 2048, 1, 188]> input_391_cast_fp16 = conv(dilations = input_391_dilations_0, groups = input_391_groups_0, pad = input_391_pad_0, pad_type = input_391_pad_type_0, strides = input_391_strides_0, weight = layers_14_conv_pointwise_conv1_weight_to_fp16, x = input_389_cast_fp16)[name = tensor<string, []>("input_391_cast_fp16")];
+            tensor<int32, []> input_393_split_num_splits_0 = const()[name = tensor<string, []>("input_393_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> input_393_split_axis_0 = const()[name = tensor<string, []>("input_393_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 1, 188]> input_393_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_393_split_cast_fp16_1 = split(axis = input_393_split_axis_0, num_splits = input_393_split_num_splits_0, x = input_391_cast_fp16)[name = tensor<string, []>("input_393_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_393_split_1_sigmoid_cast_fp16 = sigmoid(x = input_393_split_cast_fp16_1)[name = tensor<string, []>("input_393_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_393_cast_fp16 = mul(x = input_393_split_cast_fp16_0, y = input_393_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_393_cast_fp16")];
+            tensor<string, []> input_395_pad_type_0 = const()[name = tensor<string, []>("input_395_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_395_pad_0 = const()[name = tensor<string, []>("input_395_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
+            tensor<int32, []> input_395_groups_0 = const()[name = tensor<string, []>("input_395_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [2]> input_395_strides_0 = const()[name = tensor<string, []>("input_395_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> input_395_dilations_0 = const()[name = tensor<string, []>("input_395_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<fp16, [1024, 1, 1, 9]> const_296_to_fp16 = const()[name = tensor<string, []>("const_296_to_fp16"), val = tensor<fp16, [1024, 1, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(746221504)))];
             tensor<fp16, [1024]> const_297_to_fp16 = const()[name = tensor<string, []>("const_297_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(746240000)))];
-            tensor<fp16, [1, 1024, 1, 188]> input_395_cast_fp16 = conv(bias = const_297_to_fp16, dilations = input_393_dilations_0, groups = input_393_groups_0, pad = input_393_pad_0, pad_type = input_393_pad_type_0, strides = input_393_strides_0, weight = const_296_to_fp16, x = input_391_cast_fp16)[name = tensor<string, []>("input_395_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_397_cast_fp16 = silu(x = input_395_cast_fp16)[name = tensor<string, []>("input_397_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_397_cast_fp16 = conv(bias = const_297_to_fp16, dilations = input_395_dilations_0, groups = input_395_groups_0, pad = input_395_pad_0, pad_type = input_395_pad_type_0, strides = input_395_strides_0, weight = const_296_to_fp16, x = input_393_cast_fp16)[name = tensor<string, []>("input_397_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_399_cast_fp16 = silu(x = input_397_cast_fp16)[name = tensor<string, []>("input_399_cast_fp16")];
             tensor<string, []> x_89_pad_type_0 = const()[name = tensor<string, []>("x_89_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_89_strides_0 = const()[name = tensor<string, []>("x_89_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_89_pad_0 = const()[name = tensor<string, []>("x_89_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_89_dilations_0 = const()[name = tensor<string, []>("x_89_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_89_groups_0 = const()[name = tensor<string, []>("x_89_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_14_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("layers_14_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(746242112)))];
-            tensor<fp16, [1, 1024, 1, 188]> x_89_cast_fp16 = conv(dilations = x_89_dilations_0, groups = x_89_groups_0, pad = x_89_pad_0, pad_type = x_89_pad_type_0, strides = x_89_strides_0, weight = layers_14_conv_pointwise_conv2_weight_to_fp16, x = input_397_cast_fp16)[name = tensor<string, []>("x_89_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> x_89_cast_fp16 = conv(dilations = x_89_dilations_0, groups = x_89_groups_0, pad = x_89_pad_0, pad_type = x_89_pad_type_0, strides = x_89_strides_0, weight = layers_14_conv_pointwise_conv2_weight_to_fp16, x = input_399_cast_fp16)[name = tensor<string, []>("x_89_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_147_cast_fp16 = add(x = inputs_145_cast_fp16, y = x_89_cast_fp16)[name = tensor<string, []>("inputs_147_cast_fp16")];
             tensor<int32, [1]> out_147_axes_0 = const()[name = tensor<string, []>("out_147_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_4311_to_fp16 = const()[name = tensor<string, []>("op_4311_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_147_cast_fp16 = layer_norm(axes = out_147_axes_0, epsilon = var_4311_to_fp16, x = inputs_147_cast_fp16)[name = tensor<string, []>("out_147_cast_fp16")];
-            tensor<fp16, [1024]> input_399_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_399_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(748339328)))];
-            tensor<fp16, [1024]> input_399_beta_0_to_fp16 = const()[name = tensor<string, []>("input_399_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(748341440)))];
-            tensor<fp16, []> input_399_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_399_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_399_cast_fp16 = batch_norm(beta = input_399_beta_0_to_fp16, epsilon = input_399_epsilon_0_to_fp16, gamma = input_399_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_147_cast_fp16)[name = tensor<string, []>("input_399_cast_fp16")];
-            tensor<string, []> input_401_pad_type_0 = const()[name = tensor<string, []>("input_401_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_401_strides_0 = const()[name = tensor<string, []>("input_401_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_401_pad_0 = const()[name = tensor<string, []>("input_401_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_401_dilations_0 = const()[name = tensor<string, []>("input_401_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_401_groups_0 = const()[name = tensor<string, []>("input_401_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_4309_to_fp16 = const()[name = tensor<string, []>("op_4309_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_147_cast_fp16 = layer_norm(axes = out_147_axes_0, epsilon = var_4309_to_fp16, x = inputs_147_cast_fp16)[name = tensor<string, []>("out_147_cast_fp16")];
+            tensor<fp16, [1024]> input_401_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_401_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(748339328)))];
+            tensor<fp16, [1024]> input_401_beta_0_to_fp16 = const()[name = tensor<string, []>("input_401_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(748341440)))];
+            tensor<fp16, []> input_401_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_401_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_401_cast_fp16 = batch_norm(beta = input_401_beta_0_to_fp16, epsilon = input_401_epsilon_0_to_fp16, gamma = input_401_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_147_cast_fp16)[name = tensor<string, []>("input_401_cast_fp16")];
+            tensor<string, []> input_403_pad_type_0 = const()[name = tensor<string, []>("input_403_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_403_strides_0 = const()[name = tensor<string, []>("input_403_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_403_pad_0 = const()[name = tensor<string, []>("input_403_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_403_dilations_0 = const()[name = tensor<string, []>("input_403_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_403_groups_0 = const()[name = tensor<string, []>("input_403_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_14_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_14_feed_forward2_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(748343552)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_401_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_401_dilations_0, groups = input_401_groups_0, pad = input_401_pad_0, pad_type = input_401_pad_type_0, strides = input_401_strides_0, weight = layers_14_feed_forward2_fc1_weight_to_fp16, x = input_399_cast_fp16)[name = tensor<string, []>("input_401_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_403_cast_fp16 = silu(x = input_401_cast_fp16)[name = tensor<string, []>("input_403_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_403_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_403_dilations_0, groups = input_403_groups_0, pad = input_403_pad_0, pad_type = input_403_pad_type_0, strides = input_403_strides_0, weight = layers_14_feed_forward2_fc1_weight_to_fp16, x = input_401_cast_fp16)[name = tensor<string, []>("input_403_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_405_cast_fp16 = silu(x = input_403_cast_fp16)[name = tensor<string, []>("input_405_cast_fp16")];
             tensor<string, []> x_91_pad_type_0 = const()[name = tensor<string, []>("x_91_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_91_strides_0 = const()[name = tensor<string, []>("x_91_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_91_pad_0 = const()[name = tensor<string, []>("x_91_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_91_dilations_0 = const()[name = tensor<string, []>("x_91_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_91_groups_0 = const()[name = tensor<string, []>("x_91_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_4339_weight_0_to_fp16 = const()[name = tensor<string, []>("op_4339_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(756732224)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_4339_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_91_dilations_0, groups = x_91_groups_0, pad = x_91_pad_0, pad_type = x_91_pad_type_0, strides = x_91_strides_0, weight = var_4339_weight_0_to_fp16, x = input_403_cast_fp16)[name = tensor<string, []>("op_4339_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_149_cast_fp16 = add(x = inputs_147_cast_fp16, y = var_4339_cast_fp16)[name = tensor<string, []>("inputs_149_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_4337_weight_0_to_fp16 = const()[name = tensor<string, []>("op_4337_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(756732224)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_4337_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_91_dilations_0, groups = x_91_groups_0, pad = x_91_pad_0, pad_type = x_91_pad_type_0, strides = x_91_strides_0, weight = var_4337_weight_0_to_fp16, x = input_405_cast_fp16)[name = tensor<string, []>("op_4337_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_149_cast_fp16 = add(x = inputs_147_cast_fp16, y = var_4337_cast_fp16)[name = tensor<string, []>("inputs_149_cast_fp16")];
             tensor<int32, [1]> out_149_axes_0 = const()[name = tensor<string, []>("out_149_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_4349_to_fp16 = const()[name = tensor<string, []>("op_4349_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_149_cast_fp16 = layer_norm(axes = out_149_axes_0, epsilon = var_4349_to_fp16, x = inputs_149_cast_fp16)[name = tensor<string, []>("out_149_cast_fp16")];
+            tensor<fp16, []> var_4347_to_fp16 = const()[name = tensor<string, []>("op_4347_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_149_cast_fp16 = layer_norm(axes = out_149_axes_0, epsilon = var_4347_to_fp16, x = inputs_149_cast_fp16)[name = tensor<string, []>("out_149_cast_fp16")];
             tensor<fp16, [1024]> inputs_151_gamma_0_to_fp16 = const()[name = tensor<string, []>("inputs_151_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(765120896)))];
             tensor<fp16, [1024]> inputs_151_beta_0_to_fp16 = const()[name = tensor<string, []>("inputs_151_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(765123008)))];
             tensor<fp16, []> inputs_151_epsilon_0_to_fp16 = const()[name = tensor<string, []>("inputs_151_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_151_cast_fp16 = batch_norm(beta = inputs_151_beta_0_to_fp16, epsilon = inputs_151_epsilon_0_to_fp16, gamma = inputs_151_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_149_cast_fp16)[name = tensor<string, []>("inputs_151_cast_fp16")];
-            tensor<int32, []> var_4363 = const()[name = tensor<string, []>("op_4363"), val = tensor<int32, []>(3)];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_151_cast_fp16 = batch_norm(beta = inputs_151_beta_0_to_fp16, epsilon = inputs_151_epsilon_0_to_fp16, gamma = inputs_151_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_149_cast_fp16)[name = tensor<string, []>("inputs_151_cast_fp16")];
+            tensor<int32, []> var_4361 = const()[name = tensor<string, []>("op_4361"), val = tensor<int32, []>(3)];
             tensor<int32, [1]> out_151_axes_0 = const()[name = tensor<string, []>("out_151_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_4394_to_fp16 = const()[name = tensor<string, []>("op_4394_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_151_cast_fp16 = layer_norm(axes = out_151_axes_0, epsilon = var_4394_to_fp16, x = inputs_151_cast_fp16)[name = tensor<string, []>("out_151_cast_fp16")];
-            tensor<fp16, [1024]> input_405_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_405_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(765125120)))];
-            tensor<fp16, [1024]> input_405_beta_0_to_fp16 = const()[name = tensor<string, []>("input_405_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(765127232)))];
-            tensor<fp16, []> input_405_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_405_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_405_cast_fp16 = batch_norm(beta = input_405_beta_0_to_fp16, epsilon = input_405_epsilon_0_to_fp16, gamma = input_405_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_151_cast_fp16)[name = tensor<string, []>("input_405_cast_fp16")];
-            tensor<string, []> input_407_pad_type_0 = const()[name = tensor<string, []>("input_407_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_407_strides_0 = const()[name = tensor<string, []>("input_407_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_407_pad_0 = const()[name = tensor<string, []>("input_407_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_407_dilations_0 = const()[name = tensor<string, []>("input_407_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_407_groups_0 = const()[name = tensor<string, []>("input_407_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_4392_to_fp16 = const()[name = tensor<string, []>("op_4392_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_151_cast_fp16 = layer_norm(axes = out_151_axes_0, epsilon = var_4392_to_fp16, x = inputs_151_cast_fp16)[name = tensor<string, []>("out_151_cast_fp16")];
+            tensor<fp16, [1024]> input_407_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_407_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(765125120)))];
+            tensor<fp16, [1024]> input_407_beta_0_to_fp16 = const()[name = tensor<string, []>("input_407_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(765127232)))];
+            tensor<fp16, []> input_407_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_407_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_407_cast_fp16 = batch_norm(beta = input_407_beta_0_to_fp16, epsilon = input_407_epsilon_0_to_fp16, gamma = input_407_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_151_cast_fp16)[name = tensor<string, []>("input_407_cast_fp16")];
+            tensor<string, []> input_409_pad_type_0 = const()[name = tensor<string, []>("input_409_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_409_strides_0 = const()[name = tensor<string, []>("input_409_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_409_pad_0 = const()[name = tensor<string, []>("input_409_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_409_dilations_0 = const()[name = tensor<string, []>("input_409_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_409_groups_0 = const()[name = tensor<string, []>("input_409_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_15_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_15_feed_forward1_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(765129344)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_407_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_407_dilations_0, groups = input_407_groups_0, pad = input_407_pad_0, pad_type = input_407_pad_type_0, strides = input_407_strides_0, weight = layers_15_feed_forward1_fc1_weight_to_fp16, x = input_405_cast_fp16)[name = tensor<string, []>("input_407_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_409_cast_fp16 = silu(x = input_407_cast_fp16)[name = tensor<string, []>("input_409_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_409_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_409_dilations_0, groups = input_409_groups_0, pad = input_409_pad_0, pad_type = input_409_pad_type_0, strides = input_409_strides_0, weight = layers_15_feed_forward1_fc1_weight_to_fp16, x = input_407_cast_fp16)[name = tensor<string, []>("input_409_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_411_cast_fp16 = silu(x = input_409_cast_fp16)[name = tensor<string, []>("input_411_cast_fp16")];
             tensor<string, []> x_93_pad_type_0 = const()[name = tensor<string, []>("x_93_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_93_strides_0 = const()[name = tensor<string, []>("x_93_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_93_pad_0 = const()[name = tensor<string, []>("x_93_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_93_dilations_0 = const()[name = tensor<string, []>("x_93_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_93_groups_0 = const()[name = tensor<string, []>("x_93_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_4422_weight_0_to_fp16 = const()[name = tensor<string, []>("op_4422_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(773518016)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_4422_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_93_dilations_0, groups = x_93_groups_0, pad = x_93_pad_0, pad_type = x_93_pad_type_0, strides = x_93_strides_0, weight = var_4422_weight_0_to_fp16, x = input_409_cast_fp16)[name = tensor<string, []>("op_4422_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_153_cast_fp16 = add(x = inputs_151_cast_fp16, y = var_4422_cast_fp16)[name = tensor<string, []>("inputs_153_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_4420_weight_0_to_fp16 = const()[name = tensor<string, []>("op_4420_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(773518016)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_4420_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_93_dilations_0, groups = x_93_groups_0, pad = x_93_pad_0, pad_type = x_93_pad_type_0, strides = x_93_strides_0, weight = var_4420_weight_0_to_fp16, x = input_411_cast_fp16)[name = tensor<string, []>("op_4420_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_153_cast_fp16 = add(x = inputs_151_cast_fp16, y = var_4420_cast_fp16)[name = tensor<string, []>("inputs_153_cast_fp16")];
             tensor<int32, [1]> out_153_axes_0 = const()[name = tensor<string, []>("out_153_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_4432_to_fp16 = const()[name = tensor<string, []>("op_4432_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_153_cast_fp16 = layer_norm(axes = out_153_axes_0, epsilon = var_4432_to_fp16, x = inputs_153_cast_fp16)[name = tensor<string, []>("out_153_cast_fp16")];
+            tensor<fp16, []> var_4430_to_fp16 = const()[name = tensor<string, []>("op_4430_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_153_cast_fp16 = layer_norm(axes = out_153_axes_0, epsilon = var_4430_to_fp16, x = inputs_153_cast_fp16)[name = tensor<string, []>("out_153_cast_fp16")];
             tensor<fp16, [1024]> obj_63_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_63_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(781906688)))];
             tensor<fp16, [1024]> obj_63_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_63_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(781908800)))];
             tensor<fp16, []> obj_63_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_63_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> obj_63_cast_fp16 = batch_norm(beta = obj_63_beta_0_to_fp16, epsilon = obj_63_epsilon_0_to_fp16, gamma = obj_63_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_153_cast_fp16)[name = tensor<string, []>("obj_63_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_63_cast_fp16 = batch_norm(beta = obj_63_beta_0_to_fp16, epsilon = obj_63_epsilon_0_to_fp16, gamma = obj_63_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_153_cast_fp16)[name = tensor<string, []>("obj_63_cast_fp16")];
             tensor<string, []> query_61_pad_type_0 = const()[name = tensor<string, []>("query_61_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> query_61_strides_0 = const()[name = tensor<string, []>("query_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> query_61_pad_0 = const()[name = tensor<string, []>("query_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> query_61_dilations_0 = const()[name = tensor<string, []>("query_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> query_61_groups_0 = const()[name = tensor<string, []>("query_61_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_15_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_15_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(781910912)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_61_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = query_61_dilations_0, groups = query_61_groups_0, pad = query_61_pad_0, pad_type = query_61_pad_type_0, strides = query_61_strides_0, weight = layers_15_self_attn_q_proj_weight_to_fp16, x = obj_63_cast_fp16)[name = tensor<string, []>("query_61_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> query_61_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_61_dilations_0, groups = query_61_groups_0, pad = query_61_pad_0, pad_type = query_61_pad_type_0, strides = query_61_strides_0, weight = layers_15_self_attn_q_proj_weight_to_fp16, x = obj_63_cast_fp16)[name = tensor<string, []>("query_61_cast_fp16")];
             tensor<string, []> key_31_pad_type_0 = const()[name = tensor<string, []>("key_31_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> key_31_strides_0 = const()[name = tensor<string, []>("key_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> key_31_pad_0 = const()[name = tensor<string, []>("key_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
@@ -2812,173 +2814,173 @@ program(1.0)
             tensor<int32, [2]> value_31_dilations_0 = const()[name = tensor<string, []>("value_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> value_31_groups_0 = const()[name = tensor<string, []>("value_31_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_15_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_15_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(786105344)))];
-            tensor<fp16, [1, 1024, 1, 188]> value_31_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = value_31_dilations_0, groups = value_31_groups_0, pad = value_31_pad_0, pad_type = value_31_pad_type_0, strides = value_31_strides_0, weight = layers_15_self_attn_v_proj_weight_to_fp16, x = obj_63_cast_fp16)[name = tensor<string, []>("value_31_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_4470_to_fp16 = const()[name = tensor<string, []>("op_4470_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(788202560)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_63_cast_fp16 = add(x = query_61_cast_fp16, y = var_4470_to_fp16)[name = tensor<string, []>("query_63_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_4473_to_fp16 = const()[name = tensor<string, []>("op_4473_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(788204672)))];
-            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_31_cast_fp16 = add(x = query_61_cast_fp16, y = var_4473_to_fp16)[name = tensor<string, []>("q_with_bias_v_31_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> value_31_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_31_dilations_0, groups = value_31_groups_0, pad = value_31_pad_0, pad_type = value_31_pad_type_0, strides = value_31_strides_0, weight = layers_15_self_attn_v_proj_weight_to_fp16, x = obj_63_cast_fp16)[name = tensor<string, []>("value_31_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_4468_to_fp16 = const()[name = tensor<string, []>("op_4468_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(788202560)))];
+            tensor<fp16, [1, 1024, 1, 188]> query_63_cast_fp16 = add(x = query_61_cast_fp16, y = var_4468_to_fp16)[name = tensor<string, []>("query_63_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_4471_to_fp16 = const()[name = tensor<string, []>("op_4471_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(788204672)))];
+            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_31_cast_fp16 = add(x = query_61_cast_fp16, y = var_4471_to_fp16)[name = tensor<string, []>("q_with_bias_v_31_cast_fp16")];
             tensor<string, []> p_31_pad_type_0 = const()[name = tensor<string, []>("p_31_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> p_31_strides_0 = const()[name = tensor<string, []>("p_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> p_31_pad_0 = const()[name = tensor<string, []>("p_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> p_31_dilations_0 = const()[name = tensor<string, []>("p_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> p_31_groups_0 = const()[name = tensor<string, []>("p_31_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_15_self_attn_linear_pos_weight_to_fp16 = const()[name = tensor<string, []>("layers_15_self_attn_linear_pos_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(788206784)))];
-            tensor<fp16, [1, 1024, 1, 375]> p_31_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = p_31_dilations_0, groups = p_31_groups_0, pad = p_31_pad_0, pad_type = p_31_pad_type_0, strides = p_31_strides_0, weight = layers_15_self_attn_linear_pos_weight_to_fp16, x = pos_enc_to_fp16)[name = tensor<string, []>("p_31_cast_fp16")];
-            tensor<int32, [4]> var_4484 = const()[name = tensor<string, []>("op_4484"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_4485_cast_fp16 = reshape(shape = var_4484, x = q_with_bias_v_31_cast_fp16)[name = tensor<string, []>("op_4485_cast_fp16")];
-            tensor<int32, [4]> var_4486 = const()[name = tensor<string, []>("op_4486"), val = tensor<int32, [4]>([1, 8, 128, -1])];
-            tensor<fp16, [1, 8, 128, 375]> var_4487_cast_fp16 = reshape(shape = var_4486, x = p_31_cast_fp16)[name = tensor<string, []>("op_4487_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 375]> p_31_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_31_dilations_0, groups = p_31_groups_0, pad = p_31_pad_0, pad_type = p_31_pad_type_0, strides = p_31_strides_0, weight = layers_15_self_attn_linear_pos_weight_to_fp16, x = obj_3_cast_fp16)[name = tensor<string, []>("p_31_cast_fp16")];
+            tensor<int32, [4]> var_4482 = const()[name = tensor<string, []>("op_4482"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_4483_cast_fp16 = reshape(shape = var_4482, x = q_with_bias_v_31_cast_fp16)[name = tensor<string, []>("op_4483_cast_fp16")];
+            tensor<int32, [4]> var_4484 = const()[name = tensor<string, []>("op_4484"), val = tensor<int32, [4]>([1, 8, 128, -1])];
+            tensor<fp16, [1, 8, 128, 375]> var_4485_cast_fp16 = reshape(shape = var_4484, x = p_31_cast_fp16)[name = tensor<string, []>("op_4485_cast_fp16")];
             tensor<bool, []> matrix_bd_121_transpose_x_0 = const()[name = tensor<string, []>("matrix_bd_121_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> matrix_bd_121_transpose_y_0 = const()[name = tensor<string, []>("matrix_bd_121_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_121_cast_fp16 = matmul(transpose_x = matrix_bd_121_transpose_x_0, transpose_y = matrix_bd_121_transpose_y_0, x = var_4485_cast_fp16, y = var_4487_cast_fp16)[name = tensor<string, []>("matrix_bd_121_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_121_cast_fp16 = matmul(transpose_x = matrix_bd_121_transpose_x_0, transpose_y = matrix_bd_121_transpose_y_0, x = var_4483_cast_fp16, y = var_4485_cast_fp16)[name = tensor<string, []>("matrix_bd_121_cast_fp16")];
             tensor<int32, [8]> matrix_bd_123_pad_0 = const()[name = tensor<string, []>("matrix_bd_123_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
             tensor<string, []> matrix_bd_123_mode_0 = const()[name = tensor<string, []>("matrix_bd_123_mode_0"), val = tensor<string, []>("constant")];
             tensor<fp16, []> const_175_to_fp16 = const()[name = tensor<string, []>("const_175_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
             tensor<fp16, [1, 8, 188, 376]> matrix_bd_123_cast_fp16 = pad(constant_val = const_175_to_fp16, mode = matrix_bd_123_mode_0, pad = matrix_bd_123_pad_0, x = matrix_bd_121_cast_fp16)[name = tensor<string, []>("matrix_bd_123_cast_fp16")];
-            tensor<int32, [4]> var_4496 = const()[name = tensor<string, []>("op_4496"), val = tensor<int32, [4]>([1, 8, -1, 188])];
-            tensor<fp16, [1, 8, 376, 188]> matrix_bd_125_cast_fp16 = reshape(shape = var_4496, x = matrix_bd_123_cast_fp16)[name = tensor<string, []>("matrix_bd_125_cast_fp16")];
-            tensor<int32, [4]> var_4500_begin_0 = const()[name = tensor<string, []>("op_4500_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
-            tensor<int32, [4]> var_4500_end_0 = const()[name = tensor<string, []>("op_4500_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
-            tensor<bool, [4]> var_4500_end_mask_0 = const()[name = tensor<string, []>("op_4500_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
-            tensor<fp16, [1, 8, 375, 188]> var_4500_cast_fp16 = slice_by_index(begin = var_4500_begin_0, end = var_4500_end_0, end_mask = var_4500_end_mask_0, x = matrix_bd_125_cast_fp16)[name = tensor<string, []>("op_4500_cast_fp16")];
-            tensor<int32, [4]> var_4501 = const()[name = tensor<string, []>("op_4501"), val = tensor<int32, [4]>([1, 8, 188, 375])];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_127_cast_fp16 = reshape(shape = var_4501, x = var_4500_cast_fp16)[name = tensor<string, []>("matrix_bd_127_cast_fp16")];
-            tensor<int32, [4]> var_4506_begin_0 = const()[name = tensor<string, []>("op_4506_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [4]> var_4506_end_0 = const()[name = tensor<string, []>("op_4506_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
-            tensor<bool, [4]> var_4506_end_mask_0 = const()[name = tensor<string, []>("op_4506_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
-            tensor<fp16, [1, 8, 188, 188]> var_4506_cast_fp16 = slice_by_index(begin = var_4506_begin_0, end = var_4506_end_0, end_mask = var_4506_end_mask_0, x = matrix_bd_127_cast_fp16)[name = tensor<string, []>("op_4506_cast_fp16")];
-            tensor<fp16, []> var_4507_to_fp16 = const()[name = tensor<string, []>("op_4507_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 188, 188]> qk_mask_31_cast_fp16 = mul(x = var_4506_cast_fp16, y = var_4507_to_fp16)[name = tensor<string, []>("qk_mask_31_cast_fp16")];
-            tensor<int32, [4]> var_4511 = const()[name = tensor<string, []>("op_4511"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> mh_q_31_cast_fp16 = reshape(shape = var_4511, x = query_63_cast_fp16)[name = tensor<string, []>("mh_q_31_cast_fp16")];
-            tensor<fp16, []> var_4513_to_fp16 = const()[name = tensor<string, []>("op_4513_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 128, 188]> var_4514_cast_fp16 = mul(x = mh_q_31_cast_fp16, y = var_4513_to_fp16)[name = tensor<string, []>("op_4514_cast_fp16")];
-            tensor<int32, [4]> var_4517 = const()[name = tensor<string, []>("op_4517"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_4518_cast_fp16 = reshape(shape = var_4517, x = key_31_cast_fp16)[name = tensor<string, []>("op_4518_cast_fp16")];
+            tensor<int32, [4]> var_4494 = const()[name = tensor<string, []>("op_4494"), val = tensor<int32, [4]>([1, 8, -1, 188])];
+            tensor<fp16, [1, 8, 376, 188]> matrix_bd_125_cast_fp16 = reshape(shape = var_4494, x = matrix_bd_123_cast_fp16)[name = tensor<string, []>("matrix_bd_125_cast_fp16")];
+            tensor<int32, [4]> var_4498_begin_0 = const()[name = tensor<string, []>("op_4498_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_4498_end_0 = const()[name = tensor<string, []>("op_4498_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
+            tensor<bool, [4]> var_4498_end_mask_0 = const()[name = tensor<string, []>("op_4498_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 375, 188]> var_4498_cast_fp16 = slice_by_index(begin = var_4498_begin_0, end = var_4498_end_0, end_mask = var_4498_end_mask_0, x = matrix_bd_125_cast_fp16)[name = tensor<string, []>("op_4498_cast_fp16")];
+            tensor<int32, [4]> var_4499 = const()[name = tensor<string, []>("op_4499"), val = tensor<int32, [4]>([1, 8, 188, 375])];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_127_cast_fp16 = reshape(shape = var_4499, x = var_4498_cast_fp16)[name = tensor<string, []>("matrix_bd_127_cast_fp16")];
+            tensor<int32, [4]> var_4504_begin_0 = const()[name = tensor<string, []>("op_4504_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_4504_end_0 = const()[name = tensor<string, []>("op_4504_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
+            tensor<bool, [4]> var_4504_end_mask_0 = const()[name = tensor<string, []>("op_4504_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 188, 188]> var_4504_cast_fp16 = slice_by_index(begin = var_4504_begin_0, end = var_4504_end_0, end_mask = var_4504_end_mask_0, x = matrix_bd_127_cast_fp16)[name = tensor<string, []>("op_4504_cast_fp16")];
+            tensor<fp16, []> var_4505_to_fp16 = const()[name = tensor<string, []>("op_4505_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 188, 188]> qk_mask_31_cast_fp16 = mul(x = var_4504_cast_fp16, y = var_4505_to_fp16)[name = tensor<string, []>("qk_mask_31_cast_fp16")];
+            tensor<int32, [4]> var_4509 = const()[name = tensor<string, []>("op_4509"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> mh_q_31_cast_fp16 = reshape(shape = var_4509, x = query_63_cast_fp16)[name = tensor<string, []>("mh_q_31_cast_fp16")];
+            tensor<fp16, []> var_4511_to_fp16 = const()[name = tensor<string, []>("op_4511_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 128, 188]> var_4512_cast_fp16 = mul(x = mh_q_31_cast_fp16, y = var_4511_to_fp16)[name = tensor<string, []>("op_4512_cast_fp16")];
+            tensor<int32, [4]> var_4515 = const()[name = tensor<string, []>("op_4515"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_4516_cast_fp16 = reshape(shape = var_4515, x = key_31_cast_fp16)[name = tensor<string, []>("op_4516_cast_fp16")];
             tensor<bool, []> mh_w_61_transpose_x_0 = const()[name = tensor<string, []>("mh_w_61_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> mh_w_61_transpose_y_0 = const()[name = tensor<string, []>("mh_w_61_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 188]> mh_w_61_cast_fp16 = matmul(transpose_x = mh_w_61_transpose_x_0, transpose_y = mh_w_61_transpose_y_0, x = var_4514_cast_fp16, y = var_4518_cast_fp16)[name = tensor<string, []>("mh_w_61_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> mh_w_61_cast_fp16 = matmul(transpose_x = mh_w_61_transpose_x_0, transpose_y = mh_w_61_transpose_y_0, x = var_4512_cast_fp16, y = var_4516_cast_fp16)[name = tensor<string, []>("mh_w_61_cast_fp16")];
             tensor<fp16, [1, 8, 188, 188]> mh_w_63_cast_fp16 = add(x = mh_w_61_cast_fp16, y = qk_mask_31_cast_fp16)[name = tensor<string, []>("mh_w_63_cast_fp16")];
-            tensor<fp16, [1, 8, 188, 188]> var_4522_cast_fp16 = softmax(axis = var_4363, x = mh_w_63_cast_fp16)[name = tensor<string, []>("op_4522_cast_fp16")];
-            tensor<int32, [4]> var_4523 = const()[name = tensor<string, []>("op_4523"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_4524_cast_fp16 = reshape(shape = var_4523, x = value_31_cast_fp16)[name = tensor<string, []>("op_4524_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> var_4520_cast_fp16 = softmax(axis = var_4361, x = mh_w_63_cast_fp16)[name = tensor<string, []>("op_4520_cast_fp16")];
+            tensor<int32, [4]> var_4521 = const()[name = tensor<string, []>("op_4521"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_4522_cast_fp16 = reshape(shape = var_4521, x = value_31_cast_fp16)[name = tensor<string, []>("op_4522_cast_fp16")];
             tensor<bool, []> attn_31_transpose_x_0 = const()[name = tensor<string, []>("attn_31_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> attn_31_transpose_y_0 = const()[name = tensor<string, []>("attn_31_transpose_y_0"), val = tensor<bool, []>(true)];
-            tensor<fp16, [1, 8, 128, 188]> attn_31_cast_fp16 = matmul(transpose_x = attn_31_transpose_x_0, transpose_y = attn_31_transpose_y_0, x = var_4524_cast_fp16, y = var_4522_cast_fp16)[name = tensor<string, []>("attn_31_cast_fp16")];
-            tensor<int32, [4]> var_4527 = const()[name = tensor<string, []>("op_4527"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
-            tensor<fp16, [1, 1024, 1, 188]> input_411_cast_fp16 = reshape(shape = var_4527, x = attn_31_cast_fp16)[name = tensor<string, []>("input_411_cast_fp16")];
+            tensor<fp16, [1, 8, 128, 188]> attn_31_cast_fp16 = matmul(transpose_x = attn_31_transpose_x_0, transpose_y = attn_31_transpose_y_0, x = var_4522_cast_fp16, y = var_4520_cast_fp16)[name = tensor<string, []>("attn_31_cast_fp16")];
+            tensor<int32, [4]> var_4525 = const()[name = tensor<string, []>("op_4525"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
+            tensor<fp16, [1, 1024, 1, 188]> input_413_cast_fp16 = reshape(shape = var_4525, x = attn_31_cast_fp16)[name = tensor<string, []>("input_413_cast_fp16")];
             tensor<string, []> obj_65_pad_type_0 = const()[name = tensor<string, []>("obj_65_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> obj_65_strides_0 = const()[name = tensor<string, []>("obj_65_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> obj_65_pad_0 = const()[name = tensor<string, []>("obj_65_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> obj_65_dilations_0 = const()[name = tensor<string, []>("obj_65_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> obj_65_groups_0 = const()[name = tensor<string, []>("obj_65_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_15_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_15_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(790304000)))];
-            tensor<fp16, [1, 1024, 1, 188]> obj_65_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = obj_65_dilations_0, groups = obj_65_groups_0, pad = obj_65_pad_0, pad_type = obj_65_pad_type_0, strides = obj_65_strides_0, weight = layers_15_self_attn_o_proj_weight_to_fp16, x = input_411_cast_fp16)[name = tensor<string, []>("obj_65_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_65_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_65_dilations_0, groups = obj_65_groups_0, pad = obj_65_pad_0, pad_type = obj_65_pad_type_0, strides = obj_65_strides_0, weight = layers_15_self_attn_o_proj_weight_to_fp16, x = input_413_cast_fp16)[name = tensor<string, []>("obj_65_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_155_cast_fp16 = add(x = inputs_153_cast_fp16, y = obj_65_cast_fp16)[name = tensor<string, []>("inputs_155_cast_fp16")];
             tensor<int32, [1]> out_155_axes_0 = const()[name = tensor<string, []>("out_155_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_4545_to_fp16 = const()[name = tensor<string, []>("op_4545_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_155_cast_fp16 = layer_norm(axes = out_155_axes_0, epsilon = var_4545_to_fp16, x = inputs_155_cast_fp16)[name = tensor<string, []>("out_155_cast_fp16")];
-            tensor<fp16, [1024]> input_413_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_413_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(792401216)))];
-            tensor<fp16, [1024]> input_413_beta_0_to_fp16 = const()[name = tensor<string, []>("input_413_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(792403328)))];
-            tensor<fp16, []> input_413_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_413_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_413_cast_fp16 = batch_norm(beta = input_413_beta_0_to_fp16, epsilon = input_413_epsilon_0_to_fp16, gamma = input_413_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_155_cast_fp16)[name = tensor<string, []>("input_413_cast_fp16")];
-            tensor<string, []> input_415_pad_type_0 = const()[name = tensor<string, []>("input_415_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_415_strides_0 = const()[name = tensor<string, []>("input_415_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_415_pad_0 = const()[name = tensor<string, []>("input_415_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_415_dilations_0 = const()[name = tensor<string, []>("input_415_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_415_groups_0 = const()[name = tensor<string, []>("input_415_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_4543_to_fp16 = const()[name = tensor<string, []>("op_4543_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_155_cast_fp16 = layer_norm(axes = out_155_axes_0, epsilon = var_4543_to_fp16, x = inputs_155_cast_fp16)[name = tensor<string, []>("out_155_cast_fp16")];
+            tensor<fp16, [1024]> input_415_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_415_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(792401216)))];
+            tensor<fp16, [1024]> input_415_beta_0_to_fp16 = const()[name = tensor<string, []>("input_415_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(792403328)))];
+            tensor<fp16, []> input_415_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_415_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_415_cast_fp16 = batch_norm(beta = input_415_beta_0_to_fp16, epsilon = input_415_epsilon_0_to_fp16, gamma = input_415_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_155_cast_fp16)[name = tensor<string, []>("input_415_cast_fp16")];
+            tensor<string, []> input_417_pad_type_0 = const()[name = tensor<string, []>("input_417_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_417_strides_0 = const()[name = tensor<string, []>("input_417_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_417_pad_0 = const()[name = tensor<string, []>("input_417_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_417_dilations_0 = const()[name = tensor<string, []>("input_417_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_417_groups_0 = const()[name = tensor<string, []>("input_417_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [2048, 1024, 1, 1]> layers_15_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("layers_15_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(792405440)))];
-            tensor<fp16, [1, 2048, 1, 188]> input_415_cast_fp16 = conv(dilations = input_415_dilations_0, groups = input_415_groups_0, pad = input_415_pad_0, pad_type = input_415_pad_type_0, strides = input_415_strides_0, weight = layers_15_conv_pointwise_conv1_weight_to_fp16, x = input_413_cast_fp16)[name = tensor<string, []>("input_415_cast_fp16")];
-            tensor<int32, []> input_417_split_num_splits_0 = const()[name = tensor<string, []>("input_417_split_num_splits_0"), val = tensor<int32, []>(2)];
-            tensor<int32, []> input_417_split_axis_0 = const()[name = tensor<string, []>("input_417_split_axis_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1, 1024, 1, 188]> input_417_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_417_split_cast_fp16_1 = split(axis = input_417_split_axis_0, num_splits = input_417_split_num_splits_0, x = input_415_cast_fp16)[name = tensor<string, []>("input_417_split_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_417_split_1_sigmoid_cast_fp16 = sigmoid(x = input_417_split_cast_fp16_1)[name = tensor<string, []>("input_417_split_1_sigmoid_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_417_cast_fp16 = mul(x = input_417_split_cast_fp16_0, y = input_417_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_417_cast_fp16")];
-            tensor<string, []> input_419_pad_type_0 = const()[name = tensor<string, []>("input_419_pad_type_0"), val = tensor<string, []>("custom")];
-            tensor<int32, [4]> input_419_pad_0 = const()[name = tensor<string, []>("input_419_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
-            tensor<int32, []> input_419_groups_0 = const()[name = tensor<string, []>("input_419_groups_0"), val = tensor<int32, []>(1024)];
-            tensor<int32, [2]> input_419_strides_0 = const()[name = tensor<string, []>("input_419_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> input_419_dilations_0 = const()[name = tensor<string, []>("input_419_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [1, 2048, 1, 188]> input_417_cast_fp16 = conv(dilations = input_417_dilations_0, groups = input_417_groups_0, pad = input_417_pad_0, pad_type = input_417_pad_type_0, strides = input_417_strides_0, weight = layers_15_conv_pointwise_conv1_weight_to_fp16, x = input_415_cast_fp16)[name = tensor<string, []>("input_417_cast_fp16")];
+            tensor<int32, []> input_419_split_num_splits_0 = const()[name = tensor<string, []>("input_419_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> input_419_split_axis_0 = const()[name = tensor<string, []>("input_419_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 1, 188]> input_419_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_419_split_cast_fp16_1 = split(axis = input_419_split_axis_0, num_splits = input_419_split_num_splits_0, x = input_417_cast_fp16)[name = tensor<string, []>("input_419_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_419_split_1_sigmoid_cast_fp16 = sigmoid(x = input_419_split_cast_fp16_1)[name = tensor<string, []>("input_419_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_419_cast_fp16 = mul(x = input_419_split_cast_fp16_0, y = input_419_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_419_cast_fp16")];
+            tensor<string, []> input_421_pad_type_0 = const()[name = tensor<string, []>("input_421_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_421_pad_0 = const()[name = tensor<string, []>("input_421_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
+            tensor<int32, []> input_421_groups_0 = const()[name = tensor<string, []>("input_421_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [2]> input_421_strides_0 = const()[name = tensor<string, []>("input_421_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> input_421_dilations_0 = const()[name = tensor<string, []>("input_421_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<fp16, [1024, 1, 1, 9]> const_298_to_fp16 = const()[name = tensor<string, []>("const_298_to_fp16"), val = tensor<fp16, [1024, 1, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(796599808)))];
             tensor<fp16, [1024]> const_299_to_fp16 = const()[name = tensor<string, []>("const_299_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(796618304)))];
-            tensor<fp16, [1, 1024, 1, 188]> input_421_cast_fp16 = conv(bias = const_299_to_fp16, dilations = input_419_dilations_0, groups = input_419_groups_0, pad = input_419_pad_0, pad_type = input_419_pad_type_0, strides = input_419_strides_0, weight = const_298_to_fp16, x = input_417_cast_fp16)[name = tensor<string, []>("input_421_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_423_cast_fp16 = silu(x = input_421_cast_fp16)[name = tensor<string, []>("input_423_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_423_cast_fp16 = conv(bias = const_299_to_fp16, dilations = input_421_dilations_0, groups = input_421_groups_0, pad = input_421_pad_0, pad_type = input_421_pad_type_0, strides = input_421_strides_0, weight = const_298_to_fp16, x = input_419_cast_fp16)[name = tensor<string, []>("input_423_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_425_cast_fp16 = silu(x = input_423_cast_fp16)[name = tensor<string, []>("input_425_cast_fp16")];
             tensor<string, []> x_95_pad_type_0 = const()[name = tensor<string, []>("x_95_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_95_strides_0 = const()[name = tensor<string, []>("x_95_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_95_pad_0 = const()[name = tensor<string, []>("x_95_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_95_dilations_0 = const()[name = tensor<string, []>("x_95_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_95_groups_0 = const()[name = tensor<string, []>("x_95_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_15_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("layers_15_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(796620416)))];
-            tensor<fp16, [1, 1024, 1, 188]> x_95_cast_fp16 = conv(dilations = x_95_dilations_0, groups = x_95_groups_0, pad = x_95_pad_0, pad_type = x_95_pad_type_0, strides = x_95_strides_0, weight = layers_15_conv_pointwise_conv2_weight_to_fp16, x = input_423_cast_fp16)[name = tensor<string, []>("x_95_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> x_95_cast_fp16 = conv(dilations = x_95_dilations_0, groups = x_95_groups_0, pad = x_95_pad_0, pad_type = x_95_pad_type_0, strides = x_95_strides_0, weight = layers_15_conv_pointwise_conv2_weight_to_fp16, x = input_425_cast_fp16)[name = tensor<string, []>("x_95_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_157_cast_fp16 = add(x = inputs_155_cast_fp16, y = x_95_cast_fp16)[name = tensor<string, []>("inputs_157_cast_fp16")];
             tensor<int32, [1]> out_157_axes_0 = const()[name = tensor<string, []>("out_157_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_4593_to_fp16 = const()[name = tensor<string, []>("op_4593_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_157_cast_fp16 = layer_norm(axes = out_157_axes_0, epsilon = var_4593_to_fp16, x = inputs_157_cast_fp16)[name = tensor<string, []>("out_157_cast_fp16")];
-            tensor<fp16, [1024]> input_425_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_425_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(798717632)))];
-            tensor<fp16, [1024]> input_425_beta_0_to_fp16 = const()[name = tensor<string, []>("input_425_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(798719744)))];
-            tensor<fp16, []> input_425_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_425_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_425_cast_fp16 = batch_norm(beta = input_425_beta_0_to_fp16, epsilon = input_425_epsilon_0_to_fp16, gamma = input_425_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_157_cast_fp16)[name = tensor<string, []>("input_425_cast_fp16")];
-            tensor<string, []> input_427_pad_type_0 = const()[name = tensor<string, []>("input_427_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_427_strides_0 = const()[name = tensor<string, []>("input_427_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_427_pad_0 = const()[name = tensor<string, []>("input_427_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_427_dilations_0 = const()[name = tensor<string, []>("input_427_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_427_groups_0 = const()[name = tensor<string, []>("input_427_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_4591_to_fp16 = const()[name = tensor<string, []>("op_4591_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_157_cast_fp16 = layer_norm(axes = out_157_axes_0, epsilon = var_4591_to_fp16, x = inputs_157_cast_fp16)[name = tensor<string, []>("out_157_cast_fp16")];
+            tensor<fp16, [1024]> input_427_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_427_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(798717632)))];
+            tensor<fp16, [1024]> input_427_beta_0_to_fp16 = const()[name = tensor<string, []>("input_427_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(798719744)))];
+            tensor<fp16, []> input_427_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_427_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_427_cast_fp16 = batch_norm(beta = input_427_beta_0_to_fp16, epsilon = input_427_epsilon_0_to_fp16, gamma = input_427_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_157_cast_fp16)[name = tensor<string, []>("input_427_cast_fp16")];
+            tensor<string, []> input_429_pad_type_0 = const()[name = tensor<string, []>("input_429_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_429_strides_0 = const()[name = tensor<string, []>("input_429_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_429_pad_0 = const()[name = tensor<string, []>("input_429_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_429_dilations_0 = const()[name = tensor<string, []>("input_429_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_429_groups_0 = const()[name = tensor<string, []>("input_429_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_15_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_15_feed_forward2_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(798721856)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_427_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_427_dilations_0, groups = input_427_groups_0, pad = input_427_pad_0, pad_type = input_427_pad_type_0, strides = input_427_strides_0, weight = layers_15_feed_forward2_fc1_weight_to_fp16, x = input_425_cast_fp16)[name = tensor<string, []>("input_427_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_429_cast_fp16 = silu(x = input_427_cast_fp16)[name = tensor<string, []>("input_429_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_429_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_429_dilations_0, groups = input_429_groups_0, pad = input_429_pad_0, pad_type = input_429_pad_type_0, strides = input_429_strides_0, weight = layers_15_feed_forward2_fc1_weight_to_fp16, x = input_427_cast_fp16)[name = tensor<string, []>("input_429_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_431_cast_fp16 = silu(x = input_429_cast_fp16)[name = tensor<string, []>("input_431_cast_fp16")];
             tensor<string, []> x_97_pad_type_0 = const()[name = tensor<string, []>("x_97_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_97_strides_0 = const()[name = tensor<string, []>("x_97_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_97_pad_0 = const()[name = tensor<string, []>("x_97_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_97_dilations_0 = const()[name = tensor<string, []>("x_97_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_97_groups_0 = const()[name = tensor<string, []>("x_97_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_4621_weight_0_to_fp16 = const()[name = tensor<string, []>("op_4621_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(807110528)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_4621_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_97_dilations_0, groups = x_97_groups_0, pad = x_97_pad_0, pad_type = x_97_pad_type_0, strides = x_97_strides_0, weight = var_4621_weight_0_to_fp16, x = input_429_cast_fp16)[name = tensor<string, []>("op_4621_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_159_cast_fp16 = add(x = inputs_157_cast_fp16, y = var_4621_cast_fp16)[name = tensor<string, []>("inputs_159_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_4619_weight_0_to_fp16 = const()[name = tensor<string, []>("op_4619_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(807110528)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_4619_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_97_dilations_0, groups = x_97_groups_0, pad = x_97_pad_0, pad_type = x_97_pad_type_0, strides = x_97_strides_0, weight = var_4619_weight_0_to_fp16, x = input_431_cast_fp16)[name = tensor<string, []>("op_4619_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_159_cast_fp16 = add(x = inputs_157_cast_fp16, y = var_4619_cast_fp16)[name = tensor<string, []>("inputs_159_cast_fp16")];
             tensor<int32, [1]> out_159_axes_0 = const()[name = tensor<string, []>("out_159_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_4631_to_fp16 = const()[name = tensor<string, []>("op_4631_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_159_cast_fp16 = layer_norm(axes = out_159_axes_0, epsilon = var_4631_to_fp16, x = inputs_159_cast_fp16)[name = tensor<string, []>("out_159_cast_fp16")];
+            tensor<fp16, []> var_4629_to_fp16 = const()[name = tensor<string, []>("op_4629_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_159_cast_fp16 = layer_norm(axes = out_159_axes_0, epsilon = var_4629_to_fp16, x = inputs_159_cast_fp16)[name = tensor<string, []>("out_159_cast_fp16")];
             tensor<fp16, [1024]> inputs_161_gamma_0_to_fp16 = const()[name = tensor<string, []>("inputs_161_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(815499200)))];
             tensor<fp16, [1024]> inputs_161_beta_0_to_fp16 = const()[name = tensor<string, []>("inputs_161_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(815501312)))];
             tensor<fp16, []> inputs_161_epsilon_0_to_fp16 = const()[name = tensor<string, []>("inputs_161_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_161_cast_fp16 = batch_norm(beta = inputs_161_beta_0_to_fp16, epsilon = inputs_161_epsilon_0_to_fp16, gamma = inputs_161_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_159_cast_fp16)[name = tensor<string, []>("inputs_161_cast_fp16")];
-            tensor<int32, []> var_4645 = const()[name = tensor<string, []>("op_4645"), val = tensor<int32, []>(3)];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_161_cast_fp16 = batch_norm(beta = inputs_161_beta_0_to_fp16, epsilon = inputs_161_epsilon_0_to_fp16, gamma = inputs_161_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_159_cast_fp16)[name = tensor<string, []>("inputs_161_cast_fp16")];
+            tensor<int32, []> var_4643 = const()[name = tensor<string, []>("op_4643"), val = tensor<int32, []>(3)];
             tensor<int32, [1]> out_161_axes_0 = const()[name = tensor<string, []>("out_161_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_4676_to_fp16 = const()[name = tensor<string, []>("op_4676_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_161_cast_fp16 = layer_norm(axes = out_161_axes_0, epsilon = var_4676_to_fp16, x = inputs_161_cast_fp16)[name = tensor<string, []>("out_161_cast_fp16")];
-            tensor<fp16, [1024]> input_431_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_431_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(815503424)))];
-            tensor<fp16, [1024]> input_431_beta_0_to_fp16 = const()[name = tensor<string, []>("input_431_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(815505536)))];
-            tensor<fp16, []> input_431_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_431_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_431_cast_fp16 = batch_norm(beta = input_431_beta_0_to_fp16, epsilon = input_431_epsilon_0_to_fp16, gamma = input_431_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_161_cast_fp16)[name = tensor<string, []>("input_431_cast_fp16")];
-            tensor<string, []> input_433_pad_type_0 = const()[name = tensor<string, []>("input_433_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_433_strides_0 = const()[name = tensor<string, []>("input_433_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_433_pad_0 = const()[name = tensor<string, []>("input_433_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_433_dilations_0 = const()[name = tensor<string, []>("input_433_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_433_groups_0 = const()[name = tensor<string, []>("input_433_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_4674_to_fp16 = const()[name = tensor<string, []>("op_4674_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_161_cast_fp16 = layer_norm(axes = out_161_axes_0, epsilon = var_4674_to_fp16, x = inputs_161_cast_fp16)[name = tensor<string, []>("out_161_cast_fp16")];
+            tensor<fp16, [1024]> input_433_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_433_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(815503424)))];
+            tensor<fp16, [1024]> input_433_beta_0_to_fp16 = const()[name = tensor<string, []>("input_433_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(815505536)))];
+            tensor<fp16, []> input_433_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_433_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_433_cast_fp16 = batch_norm(beta = input_433_beta_0_to_fp16, epsilon = input_433_epsilon_0_to_fp16, gamma = input_433_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_161_cast_fp16)[name = tensor<string, []>("input_433_cast_fp16")];
+            tensor<string, []> input_435_pad_type_0 = const()[name = tensor<string, []>("input_435_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_435_strides_0 = const()[name = tensor<string, []>("input_435_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_435_pad_0 = const()[name = tensor<string, []>("input_435_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_435_dilations_0 = const()[name = tensor<string, []>("input_435_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_435_groups_0 = const()[name = tensor<string, []>("input_435_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_16_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_16_feed_forward1_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(815507648)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_433_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_433_dilations_0, groups = input_433_groups_0, pad = input_433_pad_0, pad_type = input_433_pad_type_0, strides = input_433_strides_0, weight = layers_16_feed_forward1_fc1_weight_to_fp16, x = input_431_cast_fp16)[name = tensor<string, []>("input_433_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_435_cast_fp16 = silu(x = input_433_cast_fp16)[name = tensor<string, []>("input_435_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_435_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_435_dilations_0, groups = input_435_groups_0, pad = input_435_pad_0, pad_type = input_435_pad_type_0, strides = input_435_strides_0, weight = layers_16_feed_forward1_fc1_weight_to_fp16, x = input_433_cast_fp16)[name = tensor<string, []>("input_435_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_437_cast_fp16 = silu(x = input_435_cast_fp16)[name = tensor<string, []>("input_437_cast_fp16")];
             tensor<string, []> x_99_pad_type_0 = const()[name = tensor<string, []>("x_99_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_99_strides_0 = const()[name = tensor<string, []>("x_99_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_99_pad_0 = const()[name = tensor<string, []>("x_99_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_99_dilations_0 = const()[name = tensor<string, []>("x_99_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_99_groups_0 = const()[name = tensor<string, []>("x_99_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_4704_weight_0_to_fp16 = const()[name = tensor<string, []>("op_4704_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(823896320)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_4704_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_99_dilations_0, groups = x_99_groups_0, pad = x_99_pad_0, pad_type = x_99_pad_type_0, strides = x_99_strides_0, weight = var_4704_weight_0_to_fp16, x = input_435_cast_fp16)[name = tensor<string, []>("op_4704_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_163_cast_fp16 = add(x = inputs_161_cast_fp16, y = var_4704_cast_fp16)[name = tensor<string, []>("inputs_163_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_4702_weight_0_to_fp16 = const()[name = tensor<string, []>("op_4702_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(823896320)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_4702_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_99_dilations_0, groups = x_99_groups_0, pad = x_99_pad_0, pad_type = x_99_pad_type_0, strides = x_99_strides_0, weight = var_4702_weight_0_to_fp16, x = input_437_cast_fp16)[name = tensor<string, []>("op_4702_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_163_cast_fp16 = add(x = inputs_161_cast_fp16, y = var_4702_cast_fp16)[name = tensor<string, []>("inputs_163_cast_fp16")];
             tensor<int32, [1]> out_163_axes_0 = const()[name = tensor<string, []>("out_163_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_4714_to_fp16 = const()[name = tensor<string, []>("op_4714_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_163_cast_fp16 = layer_norm(axes = out_163_axes_0, epsilon = var_4714_to_fp16, x = inputs_163_cast_fp16)[name = tensor<string, []>("out_163_cast_fp16")];
+            tensor<fp16, []> var_4712_to_fp16 = const()[name = tensor<string, []>("op_4712_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_163_cast_fp16 = layer_norm(axes = out_163_axes_0, epsilon = var_4712_to_fp16, x = inputs_163_cast_fp16)[name = tensor<string, []>("out_163_cast_fp16")];
             tensor<fp16, [1024]> obj_67_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_67_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(832284992)))];
             tensor<fp16, [1024]> obj_67_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_67_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(832287104)))];
             tensor<fp16, []> obj_67_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_67_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> obj_67_cast_fp16 = batch_norm(beta = obj_67_beta_0_to_fp16, epsilon = obj_67_epsilon_0_to_fp16, gamma = obj_67_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_163_cast_fp16)[name = tensor<string, []>("obj_67_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_67_cast_fp16 = batch_norm(beta = obj_67_beta_0_to_fp16, epsilon = obj_67_epsilon_0_to_fp16, gamma = obj_67_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_163_cast_fp16)[name = tensor<string, []>("obj_67_cast_fp16")];
             tensor<string, []> query_65_pad_type_0 = const()[name = tensor<string, []>("query_65_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> query_65_strides_0 = const()[name = tensor<string, []>("query_65_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> query_65_pad_0 = const()[name = tensor<string, []>("query_65_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> query_65_dilations_0 = const()[name = tensor<string, []>("query_65_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> query_65_groups_0 = const()[name = tensor<string, []>("query_65_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_16_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_16_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(832289216)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_65_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = query_65_dilations_0, groups = query_65_groups_0, pad = query_65_pad_0, pad_type = query_65_pad_type_0, strides = query_65_strides_0, weight = layers_16_self_attn_q_proj_weight_to_fp16, x = obj_67_cast_fp16)[name = tensor<string, []>("query_65_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> query_65_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_65_dilations_0, groups = query_65_groups_0, pad = query_65_pad_0, pad_type = query_65_pad_type_0, strides = query_65_strides_0, weight = layers_16_self_attn_q_proj_weight_to_fp16, x = obj_67_cast_fp16)[name = tensor<string, []>("query_65_cast_fp16")];
             tensor<string, []> key_33_pad_type_0 = const()[name = tensor<string, []>("key_33_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> key_33_strides_0 = const()[name = tensor<string, []>("key_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> key_33_pad_0 = const()[name = tensor<string, []>("key_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
@@ -2992,173 +2994,173 @@ program(1.0)
             tensor<int32, [2]> value_33_dilations_0 = const()[name = tensor<string, []>("value_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> value_33_groups_0 = const()[name = tensor<string, []>("value_33_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_16_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_16_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(836483648)))];
-            tensor<fp16, [1, 1024, 1, 188]> value_33_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = value_33_dilations_0, groups = value_33_groups_0, pad = value_33_pad_0, pad_type = value_33_pad_type_0, strides = value_33_strides_0, weight = layers_16_self_attn_v_proj_weight_to_fp16, x = obj_67_cast_fp16)[name = tensor<string, []>("value_33_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_4752_to_fp16 = const()[name = tensor<string, []>("op_4752_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(838580864)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_67_cast_fp16 = add(x = query_65_cast_fp16, y = var_4752_to_fp16)[name = tensor<string, []>("query_67_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_4755_to_fp16 = const()[name = tensor<string, []>("op_4755_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(838582976)))];
-            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_33_cast_fp16 = add(x = query_65_cast_fp16, y = var_4755_to_fp16)[name = tensor<string, []>("q_with_bias_v_33_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> value_33_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_33_dilations_0, groups = value_33_groups_0, pad = value_33_pad_0, pad_type = value_33_pad_type_0, strides = value_33_strides_0, weight = layers_16_self_attn_v_proj_weight_to_fp16, x = obj_67_cast_fp16)[name = tensor<string, []>("value_33_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_4750_to_fp16 = const()[name = tensor<string, []>("op_4750_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(838580864)))];
+            tensor<fp16, [1, 1024, 1, 188]> query_67_cast_fp16 = add(x = query_65_cast_fp16, y = var_4750_to_fp16)[name = tensor<string, []>("query_67_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_4753_to_fp16 = const()[name = tensor<string, []>("op_4753_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(838582976)))];
+            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_33_cast_fp16 = add(x = query_65_cast_fp16, y = var_4753_to_fp16)[name = tensor<string, []>("q_with_bias_v_33_cast_fp16")];
             tensor<string, []> p_33_pad_type_0 = const()[name = tensor<string, []>("p_33_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> p_33_strides_0 = const()[name = tensor<string, []>("p_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> p_33_pad_0 = const()[name = tensor<string, []>("p_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> p_33_dilations_0 = const()[name = tensor<string, []>("p_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> p_33_groups_0 = const()[name = tensor<string, []>("p_33_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_16_self_attn_linear_pos_weight_to_fp16 = const()[name = tensor<string, []>("layers_16_self_attn_linear_pos_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(838585088)))];
-            tensor<fp16, [1, 1024, 1, 375]> p_33_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = p_33_dilations_0, groups = p_33_groups_0, pad = p_33_pad_0, pad_type = p_33_pad_type_0, strides = p_33_strides_0, weight = layers_16_self_attn_linear_pos_weight_to_fp16, x = pos_enc_to_fp16)[name = tensor<string, []>("p_33_cast_fp16")];
-            tensor<int32, [4]> var_4766 = const()[name = tensor<string, []>("op_4766"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_4767_cast_fp16 = reshape(shape = var_4766, x = q_with_bias_v_33_cast_fp16)[name = tensor<string, []>("op_4767_cast_fp16")];
-            tensor<int32, [4]> var_4768 = const()[name = tensor<string, []>("op_4768"), val = tensor<int32, [4]>([1, 8, 128, -1])];
-            tensor<fp16, [1, 8, 128, 375]> var_4769_cast_fp16 = reshape(shape = var_4768, x = p_33_cast_fp16)[name = tensor<string, []>("op_4769_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 375]> p_33_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_33_dilations_0, groups = p_33_groups_0, pad = p_33_pad_0, pad_type = p_33_pad_type_0, strides = p_33_strides_0, weight = layers_16_self_attn_linear_pos_weight_to_fp16, x = obj_3_cast_fp16)[name = tensor<string, []>("p_33_cast_fp16")];
+            tensor<int32, [4]> var_4764 = const()[name = tensor<string, []>("op_4764"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_4765_cast_fp16 = reshape(shape = var_4764, x = q_with_bias_v_33_cast_fp16)[name = tensor<string, []>("op_4765_cast_fp16")];
+            tensor<int32, [4]> var_4766 = const()[name = tensor<string, []>("op_4766"), val = tensor<int32, [4]>([1, 8, 128, -1])];
+            tensor<fp16, [1, 8, 128, 375]> var_4767_cast_fp16 = reshape(shape = var_4766, x = p_33_cast_fp16)[name = tensor<string, []>("op_4767_cast_fp16")];
             tensor<bool, []> matrix_bd_129_transpose_x_0 = const()[name = tensor<string, []>("matrix_bd_129_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> matrix_bd_129_transpose_y_0 = const()[name = tensor<string, []>("matrix_bd_129_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_129_cast_fp16 = matmul(transpose_x = matrix_bd_129_transpose_x_0, transpose_y = matrix_bd_129_transpose_y_0, x = var_4767_cast_fp16, y = var_4769_cast_fp16)[name = tensor<string, []>("matrix_bd_129_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_129_cast_fp16 = matmul(transpose_x = matrix_bd_129_transpose_x_0, transpose_y = matrix_bd_129_transpose_y_0, x = var_4765_cast_fp16, y = var_4767_cast_fp16)[name = tensor<string, []>("matrix_bd_129_cast_fp16")];
             tensor<int32, [8]> matrix_bd_131_pad_0 = const()[name = tensor<string, []>("matrix_bd_131_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
             tensor<string, []> matrix_bd_131_mode_0 = const()[name = tensor<string, []>("matrix_bd_131_mode_0"), val = tensor<string, []>("constant")];
             tensor<fp16, []> const_186_to_fp16 = const()[name = tensor<string, []>("const_186_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
             tensor<fp16, [1, 8, 188, 376]> matrix_bd_131_cast_fp16 = pad(constant_val = const_186_to_fp16, mode = matrix_bd_131_mode_0, pad = matrix_bd_131_pad_0, x = matrix_bd_129_cast_fp16)[name = tensor<string, []>("matrix_bd_131_cast_fp16")];
-            tensor<int32, [4]> var_4778 = const()[name = tensor<string, []>("op_4778"), val = tensor<int32, [4]>([1, 8, -1, 188])];
-            tensor<fp16, [1, 8, 376, 188]> matrix_bd_133_cast_fp16 = reshape(shape = var_4778, x = matrix_bd_131_cast_fp16)[name = tensor<string, []>("matrix_bd_133_cast_fp16")];
-            tensor<int32, [4]> var_4782_begin_0 = const()[name = tensor<string, []>("op_4782_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
-            tensor<int32, [4]> var_4782_end_0 = const()[name = tensor<string, []>("op_4782_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
-            tensor<bool, [4]> var_4782_end_mask_0 = const()[name = tensor<string, []>("op_4782_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
-            tensor<fp16, [1, 8, 375, 188]> var_4782_cast_fp16 = slice_by_index(begin = var_4782_begin_0, end = var_4782_end_0, end_mask = var_4782_end_mask_0, x = matrix_bd_133_cast_fp16)[name = tensor<string, []>("op_4782_cast_fp16")];
-            tensor<int32, [4]> var_4783 = const()[name = tensor<string, []>("op_4783"), val = tensor<int32, [4]>([1, 8, 188, 375])];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_135_cast_fp16 = reshape(shape = var_4783, x = var_4782_cast_fp16)[name = tensor<string, []>("matrix_bd_135_cast_fp16")];
-            tensor<int32, [4]> var_4788_begin_0 = const()[name = tensor<string, []>("op_4788_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [4]> var_4788_end_0 = const()[name = tensor<string, []>("op_4788_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
-            tensor<bool, [4]> var_4788_end_mask_0 = const()[name = tensor<string, []>("op_4788_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
-            tensor<fp16, [1, 8, 188, 188]> var_4788_cast_fp16 = slice_by_index(begin = var_4788_begin_0, end = var_4788_end_0, end_mask = var_4788_end_mask_0, x = matrix_bd_135_cast_fp16)[name = tensor<string, []>("op_4788_cast_fp16")];
-            tensor<fp16, []> var_4789_to_fp16 = const()[name = tensor<string, []>("op_4789_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 188, 188]> qk_mask_33_cast_fp16 = mul(x = var_4788_cast_fp16, y = var_4789_to_fp16)[name = tensor<string, []>("qk_mask_33_cast_fp16")];
-            tensor<int32, [4]> var_4793 = const()[name = tensor<string, []>("op_4793"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> mh_q_33_cast_fp16 = reshape(shape = var_4793, x = query_67_cast_fp16)[name = tensor<string, []>("mh_q_33_cast_fp16")];
-            tensor<fp16, []> var_4795_to_fp16 = const()[name = tensor<string, []>("op_4795_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 128, 188]> var_4796_cast_fp16 = mul(x = mh_q_33_cast_fp16, y = var_4795_to_fp16)[name = tensor<string, []>("op_4796_cast_fp16")];
-            tensor<int32, [4]> var_4799 = const()[name = tensor<string, []>("op_4799"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_4800_cast_fp16 = reshape(shape = var_4799, x = key_33_cast_fp16)[name = tensor<string, []>("op_4800_cast_fp16")];
+            tensor<int32, [4]> var_4776 = const()[name = tensor<string, []>("op_4776"), val = tensor<int32, [4]>([1, 8, -1, 188])];
+            tensor<fp16, [1, 8, 376, 188]> matrix_bd_133_cast_fp16 = reshape(shape = var_4776, x = matrix_bd_131_cast_fp16)[name = tensor<string, []>("matrix_bd_133_cast_fp16")];
+            tensor<int32, [4]> var_4780_begin_0 = const()[name = tensor<string, []>("op_4780_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_4780_end_0 = const()[name = tensor<string, []>("op_4780_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
+            tensor<bool, [4]> var_4780_end_mask_0 = const()[name = tensor<string, []>("op_4780_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 375, 188]> var_4780_cast_fp16 = slice_by_index(begin = var_4780_begin_0, end = var_4780_end_0, end_mask = var_4780_end_mask_0, x = matrix_bd_133_cast_fp16)[name = tensor<string, []>("op_4780_cast_fp16")];
+            tensor<int32, [4]> var_4781 = const()[name = tensor<string, []>("op_4781"), val = tensor<int32, [4]>([1, 8, 188, 375])];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_135_cast_fp16 = reshape(shape = var_4781, x = var_4780_cast_fp16)[name = tensor<string, []>("matrix_bd_135_cast_fp16")];
+            tensor<int32, [4]> var_4786_begin_0 = const()[name = tensor<string, []>("op_4786_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_4786_end_0 = const()[name = tensor<string, []>("op_4786_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
+            tensor<bool, [4]> var_4786_end_mask_0 = const()[name = tensor<string, []>("op_4786_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 188, 188]> var_4786_cast_fp16 = slice_by_index(begin = var_4786_begin_0, end = var_4786_end_0, end_mask = var_4786_end_mask_0, x = matrix_bd_135_cast_fp16)[name = tensor<string, []>("op_4786_cast_fp16")];
+            tensor<fp16, []> var_4787_to_fp16 = const()[name = tensor<string, []>("op_4787_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 188, 188]> qk_mask_33_cast_fp16 = mul(x = var_4786_cast_fp16, y = var_4787_to_fp16)[name = tensor<string, []>("qk_mask_33_cast_fp16")];
+            tensor<int32, [4]> var_4791 = const()[name = tensor<string, []>("op_4791"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> mh_q_33_cast_fp16 = reshape(shape = var_4791, x = query_67_cast_fp16)[name = tensor<string, []>("mh_q_33_cast_fp16")];
+            tensor<fp16, []> var_4793_to_fp16 = const()[name = tensor<string, []>("op_4793_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 128, 188]> var_4794_cast_fp16 = mul(x = mh_q_33_cast_fp16, y = var_4793_to_fp16)[name = tensor<string, []>("op_4794_cast_fp16")];
+            tensor<int32, [4]> var_4797 = const()[name = tensor<string, []>("op_4797"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_4798_cast_fp16 = reshape(shape = var_4797, x = key_33_cast_fp16)[name = tensor<string, []>("op_4798_cast_fp16")];
             tensor<bool, []> mh_w_65_transpose_x_0 = const()[name = tensor<string, []>("mh_w_65_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> mh_w_65_transpose_y_0 = const()[name = tensor<string, []>("mh_w_65_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 188]> mh_w_65_cast_fp16 = matmul(transpose_x = mh_w_65_transpose_x_0, transpose_y = mh_w_65_transpose_y_0, x = var_4796_cast_fp16, y = var_4800_cast_fp16)[name = tensor<string, []>("mh_w_65_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> mh_w_65_cast_fp16 = matmul(transpose_x = mh_w_65_transpose_x_0, transpose_y = mh_w_65_transpose_y_0, x = var_4794_cast_fp16, y = var_4798_cast_fp16)[name = tensor<string, []>("mh_w_65_cast_fp16")];
             tensor<fp16, [1, 8, 188, 188]> mh_w_67_cast_fp16 = add(x = mh_w_65_cast_fp16, y = qk_mask_33_cast_fp16)[name = tensor<string, []>("mh_w_67_cast_fp16")];
-            tensor<fp16, [1, 8, 188, 188]> var_4804_cast_fp16 = softmax(axis = var_4645, x = mh_w_67_cast_fp16)[name = tensor<string, []>("op_4804_cast_fp16")];
-            tensor<int32, [4]> var_4805 = const()[name = tensor<string, []>("op_4805"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_4806_cast_fp16 = reshape(shape = var_4805, x = value_33_cast_fp16)[name = tensor<string, []>("op_4806_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> var_4802_cast_fp16 = softmax(axis = var_4643, x = mh_w_67_cast_fp16)[name = tensor<string, []>("op_4802_cast_fp16")];
+            tensor<int32, [4]> var_4803 = const()[name = tensor<string, []>("op_4803"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_4804_cast_fp16 = reshape(shape = var_4803, x = value_33_cast_fp16)[name = tensor<string, []>("op_4804_cast_fp16")];
             tensor<bool, []> attn_33_transpose_x_0 = const()[name = tensor<string, []>("attn_33_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> attn_33_transpose_y_0 = const()[name = tensor<string, []>("attn_33_transpose_y_0"), val = tensor<bool, []>(true)];
-            tensor<fp16, [1, 8, 128, 188]> attn_33_cast_fp16 = matmul(transpose_x = attn_33_transpose_x_0, transpose_y = attn_33_transpose_y_0, x = var_4806_cast_fp16, y = var_4804_cast_fp16)[name = tensor<string, []>("attn_33_cast_fp16")];
-            tensor<int32, [4]> var_4809 = const()[name = tensor<string, []>("op_4809"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
-            tensor<fp16, [1, 1024, 1, 188]> input_437_cast_fp16 = reshape(shape = var_4809, x = attn_33_cast_fp16)[name = tensor<string, []>("input_437_cast_fp16")];
+            tensor<fp16, [1, 8, 128, 188]> attn_33_cast_fp16 = matmul(transpose_x = attn_33_transpose_x_0, transpose_y = attn_33_transpose_y_0, x = var_4804_cast_fp16, y = var_4802_cast_fp16)[name = tensor<string, []>("attn_33_cast_fp16")];
+            tensor<int32, [4]> var_4807 = const()[name = tensor<string, []>("op_4807"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
+            tensor<fp16, [1, 1024, 1, 188]> input_439_cast_fp16 = reshape(shape = var_4807, x = attn_33_cast_fp16)[name = tensor<string, []>("input_439_cast_fp16")];
             tensor<string, []> obj_69_pad_type_0 = const()[name = tensor<string, []>("obj_69_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> obj_69_strides_0 = const()[name = tensor<string, []>("obj_69_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> obj_69_pad_0 = const()[name = tensor<string, []>("obj_69_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> obj_69_dilations_0 = const()[name = tensor<string, []>("obj_69_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> obj_69_groups_0 = const()[name = tensor<string, []>("obj_69_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_16_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_16_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(840682304)))];
-            tensor<fp16, [1, 1024, 1, 188]> obj_69_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = obj_69_dilations_0, groups = obj_69_groups_0, pad = obj_69_pad_0, pad_type = obj_69_pad_type_0, strides = obj_69_strides_0, weight = layers_16_self_attn_o_proj_weight_to_fp16, x = input_437_cast_fp16)[name = tensor<string, []>("obj_69_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_69_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_69_dilations_0, groups = obj_69_groups_0, pad = obj_69_pad_0, pad_type = obj_69_pad_type_0, strides = obj_69_strides_0, weight = layers_16_self_attn_o_proj_weight_to_fp16, x = input_439_cast_fp16)[name = tensor<string, []>("obj_69_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_165_cast_fp16 = add(x = inputs_163_cast_fp16, y = obj_69_cast_fp16)[name = tensor<string, []>("inputs_165_cast_fp16")];
             tensor<int32, [1]> out_165_axes_0 = const()[name = tensor<string, []>("out_165_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_4827_to_fp16 = const()[name = tensor<string, []>("op_4827_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_165_cast_fp16 = layer_norm(axes = out_165_axes_0, epsilon = var_4827_to_fp16, x = inputs_165_cast_fp16)[name = tensor<string, []>("out_165_cast_fp16")];
-            tensor<fp16, [1024]> input_439_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_439_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(842779520)))];
-            tensor<fp16, [1024]> input_439_beta_0_to_fp16 = const()[name = tensor<string, []>("input_439_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(842781632)))];
-            tensor<fp16, []> input_439_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_439_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_439_cast_fp16 = batch_norm(beta = input_439_beta_0_to_fp16, epsilon = input_439_epsilon_0_to_fp16, gamma = input_439_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_165_cast_fp16)[name = tensor<string, []>("input_439_cast_fp16")];
-            tensor<string, []> input_441_pad_type_0 = const()[name = tensor<string, []>("input_441_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_441_strides_0 = const()[name = tensor<string, []>("input_441_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_441_pad_0 = const()[name = tensor<string, []>("input_441_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_441_dilations_0 = const()[name = tensor<string, []>("input_441_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_441_groups_0 = const()[name = tensor<string, []>("input_441_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_4825_to_fp16 = const()[name = tensor<string, []>("op_4825_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_165_cast_fp16 = layer_norm(axes = out_165_axes_0, epsilon = var_4825_to_fp16, x = inputs_165_cast_fp16)[name = tensor<string, []>("out_165_cast_fp16")];
+            tensor<fp16, [1024]> input_441_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_441_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(842779520)))];
+            tensor<fp16, [1024]> input_441_beta_0_to_fp16 = const()[name = tensor<string, []>("input_441_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(842781632)))];
+            tensor<fp16, []> input_441_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_441_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_441_cast_fp16 = batch_norm(beta = input_441_beta_0_to_fp16, epsilon = input_441_epsilon_0_to_fp16, gamma = input_441_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_165_cast_fp16)[name = tensor<string, []>("input_441_cast_fp16")];
+            tensor<string, []> input_443_pad_type_0 = const()[name = tensor<string, []>("input_443_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_443_strides_0 = const()[name = tensor<string, []>("input_443_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_443_pad_0 = const()[name = tensor<string, []>("input_443_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_443_dilations_0 = const()[name = tensor<string, []>("input_443_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_443_groups_0 = const()[name = tensor<string, []>("input_443_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [2048, 1024, 1, 1]> layers_16_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("layers_16_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(842783744)))];
-            tensor<fp16, [1, 2048, 1, 188]> input_441_cast_fp16 = conv(dilations = input_441_dilations_0, groups = input_441_groups_0, pad = input_441_pad_0, pad_type = input_441_pad_type_0, strides = input_441_strides_0, weight = layers_16_conv_pointwise_conv1_weight_to_fp16, x = input_439_cast_fp16)[name = tensor<string, []>("input_441_cast_fp16")];
-            tensor<int32, []> input_443_split_num_splits_0 = const()[name = tensor<string, []>("input_443_split_num_splits_0"), val = tensor<int32, []>(2)];
-            tensor<int32, []> input_443_split_axis_0 = const()[name = tensor<string, []>("input_443_split_axis_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1, 1024, 1, 188]> input_443_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_443_split_cast_fp16_1 = split(axis = input_443_split_axis_0, num_splits = input_443_split_num_splits_0, x = input_441_cast_fp16)[name = tensor<string, []>("input_443_split_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_443_split_1_sigmoid_cast_fp16 = sigmoid(x = input_443_split_cast_fp16_1)[name = tensor<string, []>("input_443_split_1_sigmoid_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_443_cast_fp16 = mul(x = input_443_split_cast_fp16_0, y = input_443_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_443_cast_fp16")];
-            tensor<string, []> input_445_pad_type_0 = const()[name = tensor<string, []>("input_445_pad_type_0"), val = tensor<string, []>("custom")];
-            tensor<int32, [4]> input_445_pad_0 = const()[name = tensor<string, []>("input_445_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
-            tensor<int32, []> input_445_groups_0 = const()[name = tensor<string, []>("input_445_groups_0"), val = tensor<int32, []>(1024)];
-            tensor<int32, [2]> input_445_strides_0 = const()[name = tensor<string, []>("input_445_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> input_445_dilations_0 = const()[name = tensor<string, []>("input_445_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [1, 2048, 1, 188]> input_443_cast_fp16 = conv(dilations = input_443_dilations_0, groups = input_443_groups_0, pad = input_443_pad_0, pad_type = input_443_pad_type_0, strides = input_443_strides_0, weight = layers_16_conv_pointwise_conv1_weight_to_fp16, x = input_441_cast_fp16)[name = tensor<string, []>("input_443_cast_fp16")];
+            tensor<int32, []> input_445_split_num_splits_0 = const()[name = tensor<string, []>("input_445_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> input_445_split_axis_0 = const()[name = tensor<string, []>("input_445_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 1, 188]> input_445_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_445_split_cast_fp16_1 = split(axis = input_445_split_axis_0, num_splits = input_445_split_num_splits_0, x = input_443_cast_fp16)[name = tensor<string, []>("input_445_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_445_split_1_sigmoid_cast_fp16 = sigmoid(x = input_445_split_cast_fp16_1)[name = tensor<string, []>("input_445_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_445_cast_fp16 = mul(x = input_445_split_cast_fp16_0, y = input_445_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_445_cast_fp16")];
+            tensor<string, []> input_447_pad_type_0 = const()[name = tensor<string, []>("input_447_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_447_pad_0 = const()[name = tensor<string, []>("input_447_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
+            tensor<int32, []> input_447_groups_0 = const()[name = tensor<string, []>("input_447_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [2]> input_447_strides_0 = const()[name = tensor<string, []>("input_447_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> input_447_dilations_0 = const()[name = tensor<string, []>("input_447_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<fp16, [1024, 1, 1, 9]> const_300_to_fp16 = const()[name = tensor<string, []>("const_300_to_fp16"), val = tensor<fp16, [1024, 1, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(846978112)))];
             tensor<fp16, [1024]> const_301_to_fp16 = const()[name = tensor<string, []>("const_301_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(846996608)))];
-            tensor<fp16, [1, 1024, 1, 188]> input_447_cast_fp16 = conv(bias = const_301_to_fp16, dilations = input_445_dilations_0, groups = input_445_groups_0, pad = input_445_pad_0, pad_type = input_445_pad_type_0, strides = input_445_strides_0, weight = const_300_to_fp16, x = input_443_cast_fp16)[name = tensor<string, []>("input_447_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_449_cast_fp16 = silu(x = input_447_cast_fp16)[name = tensor<string, []>("input_449_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_449_cast_fp16 = conv(bias = const_301_to_fp16, dilations = input_447_dilations_0, groups = input_447_groups_0, pad = input_447_pad_0, pad_type = input_447_pad_type_0, strides = input_447_strides_0, weight = const_300_to_fp16, x = input_445_cast_fp16)[name = tensor<string, []>("input_449_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_451_cast_fp16 = silu(x = input_449_cast_fp16)[name = tensor<string, []>("input_451_cast_fp16")];
             tensor<string, []> x_101_pad_type_0 = const()[name = tensor<string, []>("x_101_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_101_strides_0 = const()[name = tensor<string, []>("x_101_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_101_pad_0 = const()[name = tensor<string, []>("x_101_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_101_dilations_0 = const()[name = tensor<string, []>("x_101_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_101_groups_0 = const()[name = tensor<string, []>("x_101_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_16_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("layers_16_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(846998720)))];
-            tensor<fp16, [1, 1024, 1, 188]> x_101_cast_fp16 = conv(dilations = x_101_dilations_0, groups = x_101_groups_0, pad = x_101_pad_0, pad_type = x_101_pad_type_0, strides = x_101_strides_0, weight = layers_16_conv_pointwise_conv2_weight_to_fp16, x = input_449_cast_fp16)[name = tensor<string, []>("x_101_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> x_101_cast_fp16 = conv(dilations = x_101_dilations_0, groups = x_101_groups_0, pad = x_101_pad_0, pad_type = x_101_pad_type_0, strides = x_101_strides_0, weight = layers_16_conv_pointwise_conv2_weight_to_fp16, x = input_451_cast_fp16)[name = tensor<string, []>("x_101_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_167_cast_fp16 = add(x = inputs_165_cast_fp16, y = x_101_cast_fp16)[name = tensor<string, []>("inputs_167_cast_fp16")];
             tensor<int32, [1]> out_167_axes_0 = const()[name = tensor<string, []>("out_167_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_4875_to_fp16 = const()[name = tensor<string, []>("op_4875_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_167_cast_fp16 = layer_norm(axes = out_167_axes_0, epsilon = var_4875_to_fp16, x = inputs_167_cast_fp16)[name = tensor<string, []>("out_167_cast_fp16")];
-            tensor<fp16, [1024]> input_451_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_451_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(849095936)))];
-            tensor<fp16, [1024]> input_451_beta_0_to_fp16 = const()[name = tensor<string, []>("input_451_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(849098048)))];
-            tensor<fp16, []> input_451_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_451_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_451_cast_fp16 = batch_norm(beta = input_451_beta_0_to_fp16, epsilon = input_451_epsilon_0_to_fp16, gamma = input_451_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_167_cast_fp16)[name = tensor<string, []>("input_451_cast_fp16")];
-            tensor<string, []> input_453_pad_type_0 = const()[name = tensor<string, []>("input_453_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_453_strides_0 = const()[name = tensor<string, []>("input_453_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_453_pad_0 = const()[name = tensor<string, []>("input_453_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_453_dilations_0 = const()[name = tensor<string, []>("input_453_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_453_groups_0 = const()[name = tensor<string, []>("input_453_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_4873_to_fp16 = const()[name = tensor<string, []>("op_4873_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_167_cast_fp16 = layer_norm(axes = out_167_axes_0, epsilon = var_4873_to_fp16, x = inputs_167_cast_fp16)[name = tensor<string, []>("out_167_cast_fp16")];
+            tensor<fp16, [1024]> input_453_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_453_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(849095936)))];
+            tensor<fp16, [1024]> input_453_beta_0_to_fp16 = const()[name = tensor<string, []>("input_453_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(849098048)))];
+            tensor<fp16, []> input_453_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_453_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_453_cast_fp16 = batch_norm(beta = input_453_beta_0_to_fp16, epsilon = input_453_epsilon_0_to_fp16, gamma = input_453_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_167_cast_fp16)[name = tensor<string, []>("input_453_cast_fp16")];
+            tensor<string, []> input_455_pad_type_0 = const()[name = tensor<string, []>("input_455_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_455_strides_0 = const()[name = tensor<string, []>("input_455_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_455_pad_0 = const()[name = tensor<string, []>("input_455_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_455_dilations_0 = const()[name = tensor<string, []>("input_455_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_455_groups_0 = const()[name = tensor<string, []>("input_455_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_16_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_16_feed_forward2_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(849100160)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_453_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_453_dilations_0, groups = input_453_groups_0, pad = input_453_pad_0, pad_type = input_453_pad_type_0, strides = input_453_strides_0, weight = layers_16_feed_forward2_fc1_weight_to_fp16, x = input_451_cast_fp16)[name = tensor<string, []>("input_453_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_455_cast_fp16 = silu(x = input_453_cast_fp16)[name = tensor<string, []>("input_455_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_455_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_455_dilations_0, groups = input_455_groups_0, pad = input_455_pad_0, pad_type = input_455_pad_type_0, strides = input_455_strides_0, weight = layers_16_feed_forward2_fc1_weight_to_fp16, x = input_453_cast_fp16)[name = tensor<string, []>("input_455_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_457_cast_fp16 = silu(x = input_455_cast_fp16)[name = tensor<string, []>("input_457_cast_fp16")];
             tensor<string, []> x_103_pad_type_0 = const()[name = tensor<string, []>("x_103_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_103_strides_0 = const()[name = tensor<string, []>("x_103_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_103_pad_0 = const()[name = tensor<string, []>("x_103_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_103_dilations_0 = const()[name = tensor<string, []>("x_103_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_103_groups_0 = const()[name = tensor<string, []>("x_103_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_4903_weight_0_to_fp16 = const()[name = tensor<string, []>("op_4903_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(857488832)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_4903_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_103_dilations_0, groups = x_103_groups_0, pad = x_103_pad_0, pad_type = x_103_pad_type_0, strides = x_103_strides_0, weight = var_4903_weight_0_to_fp16, x = input_455_cast_fp16)[name = tensor<string, []>("op_4903_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_169_cast_fp16 = add(x = inputs_167_cast_fp16, y = var_4903_cast_fp16)[name = tensor<string, []>("inputs_169_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_4901_weight_0_to_fp16 = const()[name = tensor<string, []>("op_4901_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(857488832)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_4901_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_103_dilations_0, groups = x_103_groups_0, pad = x_103_pad_0, pad_type = x_103_pad_type_0, strides = x_103_strides_0, weight = var_4901_weight_0_to_fp16, x = input_457_cast_fp16)[name = tensor<string, []>("op_4901_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_169_cast_fp16 = add(x = inputs_167_cast_fp16, y = var_4901_cast_fp16)[name = tensor<string, []>("inputs_169_cast_fp16")];
             tensor<int32, [1]> out_169_axes_0 = const()[name = tensor<string, []>("out_169_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_4913_to_fp16 = const()[name = tensor<string, []>("op_4913_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_169_cast_fp16 = layer_norm(axes = out_169_axes_0, epsilon = var_4913_to_fp16, x = inputs_169_cast_fp16)[name = tensor<string, []>("out_169_cast_fp16")];
+            tensor<fp16, []> var_4911_to_fp16 = const()[name = tensor<string, []>("op_4911_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_169_cast_fp16 = layer_norm(axes = out_169_axes_0, epsilon = var_4911_to_fp16, x = inputs_169_cast_fp16)[name = tensor<string, []>("out_169_cast_fp16")];
             tensor<fp16, [1024]> inputs_171_gamma_0_to_fp16 = const()[name = tensor<string, []>("inputs_171_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(865877504)))];
             tensor<fp16, [1024]> inputs_171_beta_0_to_fp16 = const()[name = tensor<string, []>("inputs_171_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(865879616)))];
             tensor<fp16, []> inputs_171_epsilon_0_to_fp16 = const()[name = tensor<string, []>("inputs_171_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_171_cast_fp16 = batch_norm(beta = inputs_171_beta_0_to_fp16, epsilon = inputs_171_epsilon_0_to_fp16, gamma = inputs_171_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_169_cast_fp16)[name = tensor<string, []>("inputs_171_cast_fp16")];
-            tensor<int32, []> var_4927 = const()[name = tensor<string, []>("op_4927"), val = tensor<int32, []>(3)];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_171_cast_fp16 = batch_norm(beta = inputs_171_beta_0_to_fp16, epsilon = inputs_171_epsilon_0_to_fp16, gamma = inputs_171_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_169_cast_fp16)[name = tensor<string, []>("inputs_171_cast_fp16")];
+            tensor<int32, []> var_4925 = const()[name = tensor<string, []>("op_4925"), val = tensor<int32, []>(3)];
             tensor<int32, [1]> out_171_axes_0 = const()[name = tensor<string, []>("out_171_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_4958_to_fp16 = const()[name = tensor<string, []>("op_4958_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_171_cast_fp16 = layer_norm(axes = out_171_axes_0, epsilon = var_4958_to_fp16, x = inputs_171_cast_fp16)[name = tensor<string, []>("out_171_cast_fp16")];
-            tensor<fp16, [1024]> input_457_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_457_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(865881728)))];
-            tensor<fp16, [1024]> input_457_beta_0_to_fp16 = const()[name = tensor<string, []>("input_457_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(865883840)))];
-            tensor<fp16, []> input_457_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_457_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_457_cast_fp16 = batch_norm(beta = input_457_beta_0_to_fp16, epsilon = input_457_epsilon_0_to_fp16, gamma = input_457_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_171_cast_fp16)[name = tensor<string, []>("input_457_cast_fp16")];
-            tensor<string, []> input_459_pad_type_0 = const()[name = tensor<string, []>("input_459_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_459_strides_0 = const()[name = tensor<string, []>("input_459_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_459_pad_0 = const()[name = tensor<string, []>("input_459_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_459_dilations_0 = const()[name = tensor<string, []>("input_459_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_459_groups_0 = const()[name = tensor<string, []>("input_459_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_4956_to_fp16 = const()[name = tensor<string, []>("op_4956_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_171_cast_fp16 = layer_norm(axes = out_171_axes_0, epsilon = var_4956_to_fp16, x = inputs_171_cast_fp16)[name = tensor<string, []>("out_171_cast_fp16")];
+            tensor<fp16, [1024]> input_459_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_459_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(865881728)))];
+            tensor<fp16, [1024]> input_459_beta_0_to_fp16 = const()[name = tensor<string, []>("input_459_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(865883840)))];
+            tensor<fp16, []> input_459_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_459_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_459_cast_fp16 = batch_norm(beta = input_459_beta_0_to_fp16, epsilon = input_459_epsilon_0_to_fp16, gamma = input_459_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_171_cast_fp16)[name = tensor<string, []>("input_459_cast_fp16")];
+            tensor<string, []> input_461_pad_type_0 = const()[name = tensor<string, []>("input_461_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_461_strides_0 = const()[name = tensor<string, []>("input_461_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_461_pad_0 = const()[name = tensor<string, []>("input_461_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_461_dilations_0 = const()[name = tensor<string, []>("input_461_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_461_groups_0 = const()[name = tensor<string, []>("input_461_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_17_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_17_feed_forward1_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(865885952)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_459_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_459_dilations_0, groups = input_459_groups_0, pad = input_459_pad_0, pad_type = input_459_pad_type_0, strides = input_459_strides_0, weight = layers_17_feed_forward1_fc1_weight_to_fp16, x = input_457_cast_fp16)[name = tensor<string, []>("input_459_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_461_cast_fp16 = silu(x = input_459_cast_fp16)[name = tensor<string, []>("input_461_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_461_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_461_dilations_0, groups = input_461_groups_0, pad = input_461_pad_0, pad_type = input_461_pad_type_0, strides = input_461_strides_0, weight = layers_17_feed_forward1_fc1_weight_to_fp16, x = input_459_cast_fp16)[name = tensor<string, []>("input_461_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_463_cast_fp16 = silu(x = input_461_cast_fp16)[name = tensor<string, []>("input_463_cast_fp16")];
             tensor<string, []> x_105_pad_type_0 = const()[name = tensor<string, []>("x_105_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_105_strides_0 = const()[name = tensor<string, []>("x_105_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_105_pad_0 = const()[name = tensor<string, []>("x_105_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_105_dilations_0 = const()[name = tensor<string, []>("x_105_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_105_groups_0 = const()[name = tensor<string, []>("x_105_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_4986_weight_0_to_fp16 = const()[name = tensor<string, []>("op_4986_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(874274624)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_4986_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_105_dilations_0, groups = x_105_groups_0, pad = x_105_pad_0, pad_type = x_105_pad_type_0, strides = x_105_strides_0, weight = var_4986_weight_0_to_fp16, x = input_461_cast_fp16)[name = tensor<string, []>("op_4986_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_173_cast_fp16 = add(x = inputs_171_cast_fp16, y = var_4986_cast_fp16)[name = tensor<string, []>("inputs_173_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_4984_weight_0_to_fp16 = const()[name = tensor<string, []>("op_4984_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(874274624)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_4984_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_105_dilations_0, groups = x_105_groups_0, pad = x_105_pad_0, pad_type = x_105_pad_type_0, strides = x_105_strides_0, weight = var_4984_weight_0_to_fp16, x = input_463_cast_fp16)[name = tensor<string, []>("op_4984_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_173_cast_fp16 = add(x = inputs_171_cast_fp16, y = var_4984_cast_fp16)[name = tensor<string, []>("inputs_173_cast_fp16")];
             tensor<int32, [1]> out_173_axes_0 = const()[name = tensor<string, []>("out_173_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_4996_to_fp16 = const()[name = tensor<string, []>("op_4996_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_173_cast_fp16 = layer_norm(axes = out_173_axes_0, epsilon = var_4996_to_fp16, x = inputs_173_cast_fp16)[name = tensor<string, []>("out_173_cast_fp16")];
+            tensor<fp16, []> var_4994_to_fp16 = const()[name = tensor<string, []>("op_4994_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_173_cast_fp16 = layer_norm(axes = out_173_axes_0, epsilon = var_4994_to_fp16, x = inputs_173_cast_fp16)[name = tensor<string, []>("out_173_cast_fp16")];
             tensor<fp16, [1024]> obj_71_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_71_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(882663296)))];
             tensor<fp16, [1024]> obj_71_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_71_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(882665408)))];
             tensor<fp16, []> obj_71_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_71_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> obj_71_cast_fp16 = batch_norm(beta = obj_71_beta_0_to_fp16, epsilon = obj_71_epsilon_0_to_fp16, gamma = obj_71_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_173_cast_fp16)[name = tensor<string, []>("obj_71_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_71_cast_fp16 = batch_norm(beta = obj_71_beta_0_to_fp16, epsilon = obj_71_epsilon_0_to_fp16, gamma = obj_71_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_173_cast_fp16)[name = tensor<string, []>("obj_71_cast_fp16")];
             tensor<string, []> query_69_pad_type_0 = const()[name = tensor<string, []>("query_69_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> query_69_strides_0 = const()[name = tensor<string, []>("query_69_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> query_69_pad_0 = const()[name = tensor<string, []>("query_69_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> query_69_dilations_0 = const()[name = tensor<string, []>("query_69_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> query_69_groups_0 = const()[name = tensor<string, []>("query_69_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_17_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_17_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(882667520)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_69_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = query_69_dilations_0, groups = query_69_groups_0, pad = query_69_pad_0, pad_type = query_69_pad_type_0, strides = query_69_strides_0, weight = layers_17_self_attn_q_proj_weight_to_fp16, x = obj_71_cast_fp16)[name = tensor<string, []>("query_69_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> query_69_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_69_dilations_0, groups = query_69_groups_0, pad = query_69_pad_0, pad_type = query_69_pad_type_0, strides = query_69_strides_0, weight = layers_17_self_attn_q_proj_weight_to_fp16, x = obj_71_cast_fp16)[name = tensor<string, []>("query_69_cast_fp16")];
             tensor<string, []> key_35_pad_type_0 = const()[name = tensor<string, []>("key_35_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> key_35_strides_0 = const()[name = tensor<string, []>("key_35_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> key_35_pad_0 = const()[name = tensor<string, []>("key_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
@@ -3172,173 +3174,173 @@ program(1.0)
             tensor<int32, [2]> value_35_dilations_0 = const()[name = tensor<string, []>("value_35_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> value_35_groups_0 = const()[name = tensor<string, []>("value_35_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_17_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_17_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(886861952)))];
-            tensor<fp16, [1, 1024, 1, 188]> value_35_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = value_35_dilations_0, groups = value_35_groups_0, pad = value_35_pad_0, pad_type = value_35_pad_type_0, strides = value_35_strides_0, weight = layers_17_self_attn_v_proj_weight_to_fp16, x = obj_71_cast_fp16)[name = tensor<string, []>("value_35_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_5034_to_fp16 = const()[name = tensor<string, []>("op_5034_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(888959168)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_71_cast_fp16 = add(x = query_69_cast_fp16, y = var_5034_to_fp16)[name = tensor<string, []>("query_71_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_5037_to_fp16 = const()[name = tensor<string, []>("op_5037_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(888961280)))];
-            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_35_cast_fp16 = add(x = query_69_cast_fp16, y = var_5037_to_fp16)[name = tensor<string, []>("q_with_bias_v_35_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> value_35_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_35_dilations_0, groups = value_35_groups_0, pad = value_35_pad_0, pad_type = value_35_pad_type_0, strides = value_35_strides_0, weight = layers_17_self_attn_v_proj_weight_to_fp16, x = obj_71_cast_fp16)[name = tensor<string, []>("value_35_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_5032_to_fp16 = const()[name = tensor<string, []>("op_5032_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(888959168)))];
+            tensor<fp16, [1, 1024, 1, 188]> query_71_cast_fp16 = add(x = query_69_cast_fp16, y = var_5032_to_fp16)[name = tensor<string, []>("query_71_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_5035_to_fp16 = const()[name = tensor<string, []>("op_5035_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(888961280)))];
+            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_35_cast_fp16 = add(x = query_69_cast_fp16, y = var_5035_to_fp16)[name = tensor<string, []>("q_with_bias_v_35_cast_fp16")];
             tensor<string, []> p_35_pad_type_0 = const()[name = tensor<string, []>("p_35_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> p_35_strides_0 = const()[name = tensor<string, []>("p_35_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> p_35_pad_0 = const()[name = tensor<string, []>("p_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> p_35_dilations_0 = const()[name = tensor<string, []>("p_35_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> p_35_groups_0 = const()[name = tensor<string, []>("p_35_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_17_self_attn_linear_pos_weight_to_fp16 = const()[name = tensor<string, []>("layers_17_self_attn_linear_pos_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(888963392)))];
-            tensor<fp16, [1, 1024, 1, 375]> p_35_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = p_35_dilations_0, groups = p_35_groups_0, pad = p_35_pad_0, pad_type = p_35_pad_type_0, strides = p_35_strides_0, weight = layers_17_self_attn_linear_pos_weight_to_fp16, x = pos_enc_to_fp16)[name = tensor<string, []>("p_35_cast_fp16")];
-            tensor<int32, [4]> var_5048 = const()[name = tensor<string, []>("op_5048"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_5049_cast_fp16 = reshape(shape = var_5048, x = q_with_bias_v_35_cast_fp16)[name = tensor<string, []>("op_5049_cast_fp16")];
-            tensor<int32, [4]> var_5050 = const()[name = tensor<string, []>("op_5050"), val = tensor<int32, [4]>([1, 8, 128, -1])];
-            tensor<fp16, [1, 8, 128, 375]> var_5051_cast_fp16 = reshape(shape = var_5050, x = p_35_cast_fp16)[name = tensor<string, []>("op_5051_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 375]> p_35_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_35_dilations_0, groups = p_35_groups_0, pad = p_35_pad_0, pad_type = p_35_pad_type_0, strides = p_35_strides_0, weight = layers_17_self_attn_linear_pos_weight_to_fp16, x = obj_3_cast_fp16)[name = tensor<string, []>("p_35_cast_fp16")];
+            tensor<int32, [4]> var_5046 = const()[name = tensor<string, []>("op_5046"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_5047_cast_fp16 = reshape(shape = var_5046, x = q_with_bias_v_35_cast_fp16)[name = tensor<string, []>("op_5047_cast_fp16")];
+            tensor<int32, [4]> var_5048 = const()[name = tensor<string, []>("op_5048"), val = tensor<int32, [4]>([1, 8, 128, -1])];
+            tensor<fp16, [1, 8, 128, 375]> var_5049_cast_fp16 = reshape(shape = var_5048, x = p_35_cast_fp16)[name = tensor<string, []>("op_5049_cast_fp16")];
             tensor<bool, []> matrix_bd_137_transpose_x_0 = const()[name = tensor<string, []>("matrix_bd_137_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> matrix_bd_137_transpose_y_0 = const()[name = tensor<string, []>("matrix_bd_137_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_137_cast_fp16 = matmul(transpose_x = matrix_bd_137_transpose_x_0, transpose_y = matrix_bd_137_transpose_y_0, x = var_5049_cast_fp16, y = var_5051_cast_fp16)[name = tensor<string, []>("matrix_bd_137_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_137_cast_fp16 = matmul(transpose_x = matrix_bd_137_transpose_x_0, transpose_y = matrix_bd_137_transpose_y_0, x = var_5047_cast_fp16, y = var_5049_cast_fp16)[name = tensor<string, []>("matrix_bd_137_cast_fp16")];
             tensor<int32, [8]> matrix_bd_139_pad_0 = const()[name = tensor<string, []>("matrix_bd_139_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
             tensor<string, []> matrix_bd_139_mode_0 = const()[name = tensor<string, []>("matrix_bd_139_mode_0"), val = tensor<string, []>("constant")];
             tensor<fp16, []> const_197_to_fp16 = const()[name = tensor<string, []>("const_197_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
             tensor<fp16, [1, 8, 188, 376]> matrix_bd_139_cast_fp16 = pad(constant_val = const_197_to_fp16, mode = matrix_bd_139_mode_0, pad = matrix_bd_139_pad_0, x = matrix_bd_137_cast_fp16)[name = tensor<string, []>("matrix_bd_139_cast_fp16")];
-            tensor<int32, [4]> var_5060 = const()[name = tensor<string, []>("op_5060"), val = tensor<int32, [4]>([1, 8, -1, 188])];
-            tensor<fp16, [1, 8, 376, 188]> matrix_bd_141_cast_fp16 = reshape(shape = var_5060, x = matrix_bd_139_cast_fp16)[name = tensor<string, []>("matrix_bd_141_cast_fp16")];
-            tensor<int32, [4]> var_5064_begin_0 = const()[name = tensor<string, []>("op_5064_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
-            tensor<int32, [4]> var_5064_end_0 = const()[name = tensor<string, []>("op_5064_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
-            tensor<bool, [4]> var_5064_end_mask_0 = const()[name = tensor<string, []>("op_5064_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
-            tensor<fp16, [1, 8, 375, 188]> var_5064_cast_fp16 = slice_by_index(begin = var_5064_begin_0, end = var_5064_end_0, end_mask = var_5064_end_mask_0, x = matrix_bd_141_cast_fp16)[name = tensor<string, []>("op_5064_cast_fp16")];
-            tensor<int32, [4]> var_5065 = const()[name = tensor<string, []>("op_5065"), val = tensor<int32, [4]>([1, 8, 188, 375])];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_143_cast_fp16 = reshape(shape = var_5065, x = var_5064_cast_fp16)[name = tensor<string, []>("matrix_bd_143_cast_fp16")];
-            tensor<int32, [4]> var_5070_begin_0 = const()[name = tensor<string, []>("op_5070_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [4]> var_5070_end_0 = const()[name = tensor<string, []>("op_5070_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
-            tensor<bool, [4]> var_5070_end_mask_0 = const()[name = tensor<string, []>("op_5070_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
-            tensor<fp16, [1, 8, 188, 188]> var_5070_cast_fp16 = slice_by_index(begin = var_5070_begin_0, end = var_5070_end_0, end_mask = var_5070_end_mask_0, x = matrix_bd_143_cast_fp16)[name = tensor<string, []>("op_5070_cast_fp16")];
-            tensor<fp16, []> var_5071_to_fp16 = const()[name = tensor<string, []>("op_5071_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 188, 188]> qk_mask_35_cast_fp16 = mul(x = var_5070_cast_fp16, y = var_5071_to_fp16)[name = tensor<string, []>("qk_mask_35_cast_fp16")];
-            tensor<int32, [4]> var_5075 = const()[name = tensor<string, []>("op_5075"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> mh_q_35_cast_fp16 = reshape(shape = var_5075, x = query_71_cast_fp16)[name = tensor<string, []>("mh_q_35_cast_fp16")];
-            tensor<fp16, []> var_5077_to_fp16 = const()[name = tensor<string, []>("op_5077_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 128, 188]> var_5078_cast_fp16 = mul(x = mh_q_35_cast_fp16, y = var_5077_to_fp16)[name = tensor<string, []>("op_5078_cast_fp16")];
-            tensor<int32, [4]> var_5081 = const()[name = tensor<string, []>("op_5081"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_5082_cast_fp16 = reshape(shape = var_5081, x = key_35_cast_fp16)[name = tensor<string, []>("op_5082_cast_fp16")];
+            tensor<int32, [4]> var_5058 = const()[name = tensor<string, []>("op_5058"), val = tensor<int32, [4]>([1, 8, -1, 188])];
+            tensor<fp16, [1, 8, 376, 188]> matrix_bd_141_cast_fp16 = reshape(shape = var_5058, x = matrix_bd_139_cast_fp16)[name = tensor<string, []>("matrix_bd_141_cast_fp16")];
+            tensor<int32, [4]> var_5062_begin_0 = const()[name = tensor<string, []>("op_5062_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_5062_end_0 = const()[name = tensor<string, []>("op_5062_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
+            tensor<bool, [4]> var_5062_end_mask_0 = const()[name = tensor<string, []>("op_5062_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 375, 188]> var_5062_cast_fp16 = slice_by_index(begin = var_5062_begin_0, end = var_5062_end_0, end_mask = var_5062_end_mask_0, x = matrix_bd_141_cast_fp16)[name = tensor<string, []>("op_5062_cast_fp16")];
+            tensor<int32, [4]> var_5063 = const()[name = tensor<string, []>("op_5063"), val = tensor<int32, [4]>([1, 8, 188, 375])];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_143_cast_fp16 = reshape(shape = var_5063, x = var_5062_cast_fp16)[name = tensor<string, []>("matrix_bd_143_cast_fp16")];
+            tensor<int32, [4]> var_5068_begin_0 = const()[name = tensor<string, []>("op_5068_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_5068_end_0 = const()[name = tensor<string, []>("op_5068_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
+            tensor<bool, [4]> var_5068_end_mask_0 = const()[name = tensor<string, []>("op_5068_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 188, 188]> var_5068_cast_fp16 = slice_by_index(begin = var_5068_begin_0, end = var_5068_end_0, end_mask = var_5068_end_mask_0, x = matrix_bd_143_cast_fp16)[name = tensor<string, []>("op_5068_cast_fp16")];
+            tensor<fp16, []> var_5069_to_fp16 = const()[name = tensor<string, []>("op_5069_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 188, 188]> qk_mask_35_cast_fp16 = mul(x = var_5068_cast_fp16, y = var_5069_to_fp16)[name = tensor<string, []>("qk_mask_35_cast_fp16")];
+            tensor<int32, [4]> var_5073 = const()[name = tensor<string, []>("op_5073"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> mh_q_35_cast_fp16 = reshape(shape = var_5073, x = query_71_cast_fp16)[name = tensor<string, []>("mh_q_35_cast_fp16")];
+            tensor<fp16, []> var_5075_to_fp16 = const()[name = tensor<string, []>("op_5075_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 128, 188]> var_5076_cast_fp16 = mul(x = mh_q_35_cast_fp16, y = var_5075_to_fp16)[name = tensor<string, []>("op_5076_cast_fp16")];
+            tensor<int32, [4]> var_5079 = const()[name = tensor<string, []>("op_5079"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_5080_cast_fp16 = reshape(shape = var_5079, x = key_35_cast_fp16)[name = tensor<string, []>("op_5080_cast_fp16")];
             tensor<bool, []> mh_w_69_transpose_x_0 = const()[name = tensor<string, []>("mh_w_69_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> mh_w_69_transpose_y_0 = const()[name = tensor<string, []>("mh_w_69_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 188]> mh_w_69_cast_fp16 = matmul(transpose_x = mh_w_69_transpose_x_0, transpose_y = mh_w_69_transpose_y_0, x = var_5078_cast_fp16, y = var_5082_cast_fp16)[name = tensor<string, []>("mh_w_69_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> mh_w_69_cast_fp16 = matmul(transpose_x = mh_w_69_transpose_x_0, transpose_y = mh_w_69_transpose_y_0, x = var_5076_cast_fp16, y = var_5080_cast_fp16)[name = tensor<string, []>("mh_w_69_cast_fp16")];
             tensor<fp16, [1, 8, 188, 188]> mh_w_71_cast_fp16 = add(x = mh_w_69_cast_fp16, y = qk_mask_35_cast_fp16)[name = tensor<string, []>("mh_w_71_cast_fp16")];
-            tensor<fp16, [1, 8, 188, 188]> var_5086_cast_fp16 = softmax(axis = var_4927, x = mh_w_71_cast_fp16)[name = tensor<string, []>("op_5086_cast_fp16")];
-            tensor<int32, [4]> var_5087 = const()[name = tensor<string, []>("op_5087"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_5088_cast_fp16 = reshape(shape = var_5087, x = value_35_cast_fp16)[name = tensor<string, []>("op_5088_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> var_5084_cast_fp16 = softmax(axis = var_4925, x = mh_w_71_cast_fp16)[name = tensor<string, []>("op_5084_cast_fp16")];
+            tensor<int32, [4]> var_5085 = const()[name = tensor<string, []>("op_5085"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_5086_cast_fp16 = reshape(shape = var_5085, x = value_35_cast_fp16)[name = tensor<string, []>("op_5086_cast_fp16")];
             tensor<bool, []> attn_35_transpose_x_0 = const()[name = tensor<string, []>("attn_35_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> attn_35_transpose_y_0 = const()[name = tensor<string, []>("attn_35_transpose_y_0"), val = tensor<bool, []>(true)];
-            tensor<fp16, [1, 8, 128, 188]> attn_35_cast_fp16 = matmul(transpose_x = attn_35_transpose_x_0, transpose_y = attn_35_transpose_y_0, x = var_5088_cast_fp16, y = var_5086_cast_fp16)[name = tensor<string, []>("attn_35_cast_fp16")];
-            tensor<int32, [4]> var_5091 = const()[name = tensor<string, []>("op_5091"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
-            tensor<fp16, [1, 1024, 1, 188]> input_463_cast_fp16 = reshape(shape = var_5091, x = attn_35_cast_fp16)[name = tensor<string, []>("input_463_cast_fp16")];
+            tensor<fp16, [1, 8, 128, 188]> attn_35_cast_fp16 = matmul(transpose_x = attn_35_transpose_x_0, transpose_y = attn_35_transpose_y_0, x = var_5086_cast_fp16, y = var_5084_cast_fp16)[name = tensor<string, []>("attn_35_cast_fp16")];
+            tensor<int32, [4]> var_5089 = const()[name = tensor<string, []>("op_5089"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
+            tensor<fp16, [1, 1024, 1, 188]> input_465_cast_fp16 = reshape(shape = var_5089, x = attn_35_cast_fp16)[name = tensor<string, []>("input_465_cast_fp16")];
             tensor<string, []> obj_73_pad_type_0 = const()[name = tensor<string, []>("obj_73_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> obj_73_strides_0 = const()[name = tensor<string, []>("obj_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> obj_73_pad_0 = const()[name = tensor<string, []>("obj_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> obj_73_dilations_0 = const()[name = tensor<string, []>("obj_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> obj_73_groups_0 = const()[name = tensor<string, []>("obj_73_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_17_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_17_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(891060608)))];
-            tensor<fp16, [1, 1024, 1, 188]> obj_73_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = obj_73_dilations_0, groups = obj_73_groups_0, pad = obj_73_pad_0, pad_type = obj_73_pad_type_0, strides = obj_73_strides_0, weight = layers_17_self_attn_o_proj_weight_to_fp16, x = input_463_cast_fp16)[name = tensor<string, []>("obj_73_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_73_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_73_dilations_0, groups = obj_73_groups_0, pad = obj_73_pad_0, pad_type = obj_73_pad_type_0, strides = obj_73_strides_0, weight = layers_17_self_attn_o_proj_weight_to_fp16, x = input_465_cast_fp16)[name = tensor<string, []>("obj_73_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_175_cast_fp16 = add(x = inputs_173_cast_fp16, y = obj_73_cast_fp16)[name = tensor<string, []>("inputs_175_cast_fp16")];
             tensor<int32, [1]> out_175_axes_0 = const()[name = tensor<string, []>("out_175_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_5109_to_fp16 = const()[name = tensor<string, []>("op_5109_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_175_cast_fp16 = layer_norm(axes = out_175_axes_0, epsilon = var_5109_to_fp16, x = inputs_175_cast_fp16)[name = tensor<string, []>("out_175_cast_fp16")];
-            tensor<fp16, [1024]> input_465_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_465_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(893157824)))];
-            tensor<fp16, [1024]> input_465_beta_0_to_fp16 = const()[name = tensor<string, []>("input_465_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(893159936)))];
-            tensor<fp16, []> input_465_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_465_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_465_cast_fp16 = batch_norm(beta = input_465_beta_0_to_fp16, epsilon = input_465_epsilon_0_to_fp16, gamma = input_465_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_175_cast_fp16)[name = tensor<string, []>("input_465_cast_fp16")];
-            tensor<string, []> input_467_pad_type_0 = const()[name = tensor<string, []>("input_467_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_467_strides_0 = const()[name = tensor<string, []>("input_467_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_467_pad_0 = const()[name = tensor<string, []>("input_467_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_467_dilations_0 = const()[name = tensor<string, []>("input_467_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_467_groups_0 = const()[name = tensor<string, []>("input_467_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_5107_to_fp16 = const()[name = tensor<string, []>("op_5107_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_175_cast_fp16 = layer_norm(axes = out_175_axes_0, epsilon = var_5107_to_fp16, x = inputs_175_cast_fp16)[name = tensor<string, []>("out_175_cast_fp16")];
+            tensor<fp16, [1024]> input_467_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_467_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(893157824)))];
+            tensor<fp16, [1024]> input_467_beta_0_to_fp16 = const()[name = tensor<string, []>("input_467_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(893159936)))];
+            tensor<fp16, []> input_467_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_467_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_467_cast_fp16 = batch_norm(beta = input_467_beta_0_to_fp16, epsilon = input_467_epsilon_0_to_fp16, gamma = input_467_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_175_cast_fp16)[name = tensor<string, []>("input_467_cast_fp16")];
+            tensor<string, []> input_469_pad_type_0 = const()[name = tensor<string, []>("input_469_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_469_strides_0 = const()[name = tensor<string, []>("input_469_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_469_pad_0 = const()[name = tensor<string, []>("input_469_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_469_dilations_0 = const()[name = tensor<string, []>("input_469_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_469_groups_0 = const()[name = tensor<string, []>("input_469_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [2048, 1024, 1, 1]> layers_17_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("layers_17_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(893162048)))];
-            tensor<fp16, [1, 2048, 1, 188]> input_467_cast_fp16 = conv(dilations = input_467_dilations_0, groups = input_467_groups_0, pad = input_467_pad_0, pad_type = input_467_pad_type_0, strides = input_467_strides_0, weight = layers_17_conv_pointwise_conv1_weight_to_fp16, x = input_465_cast_fp16)[name = tensor<string, []>("input_467_cast_fp16")];
-            tensor<int32, []> input_469_split_num_splits_0 = const()[name = tensor<string, []>("input_469_split_num_splits_0"), val = tensor<int32, []>(2)];
-            tensor<int32, []> input_469_split_axis_0 = const()[name = tensor<string, []>("input_469_split_axis_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1, 1024, 1, 188]> input_469_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_469_split_cast_fp16_1 = split(axis = input_469_split_axis_0, num_splits = input_469_split_num_splits_0, x = input_467_cast_fp16)[name = tensor<string, []>("input_469_split_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_469_split_1_sigmoid_cast_fp16 = sigmoid(x = input_469_split_cast_fp16_1)[name = tensor<string, []>("input_469_split_1_sigmoid_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_469_cast_fp16 = mul(x = input_469_split_cast_fp16_0, y = input_469_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_469_cast_fp16")];
-            tensor<string, []> input_471_pad_type_0 = const()[name = tensor<string, []>("input_471_pad_type_0"), val = tensor<string, []>("custom")];
-            tensor<int32, [4]> input_471_pad_0 = const()[name = tensor<string, []>("input_471_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
-            tensor<int32, []> input_471_groups_0 = const()[name = tensor<string, []>("input_471_groups_0"), val = tensor<int32, []>(1024)];
-            tensor<int32, [2]> input_471_strides_0 = const()[name = tensor<string, []>("input_471_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> input_471_dilations_0 = const()[name = tensor<string, []>("input_471_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [1, 2048, 1, 188]> input_469_cast_fp16 = conv(dilations = input_469_dilations_0, groups = input_469_groups_0, pad = input_469_pad_0, pad_type = input_469_pad_type_0, strides = input_469_strides_0, weight = layers_17_conv_pointwise_conv1_weight_to_fp16, x = input_467_cast_fp16)[name = tensor<string, []>("input_469_cast_fp16")];
+            tensor<int32, []> input_471_split_num_splits_0 = const()[name = tensor<string, []>("input_471_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> input_471_split_axis_0 = const()[name = tensor<string, []>("input_471_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 1, 188]> input_471_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_471_split_cast_fp16_1 = split(axis = input_471_split_axis_0, num_splits = input_471_split_num_splits_0, x = input_469_cast_fp16)[name = tensor<string, []>("input_471_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_471_split_1_sigmoid_cast_fp16 = sigmoid(x = input_471_split_cast_fp16_1)[name = tensor<string, []>("input_471_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_471_cast_fp16 = mul(x = input_471_split_cast_fp16_0, y = input_471_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_471_cast_fp16")];
+            tensor<string, []> input_473_pad_type_0 = const()[name = tensor<string, []>("input_473_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_473_pad_0 = const()[name = tensor<string, []>("input_473_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
+            tensor<int32, []> input_473_groups_0 = const()[name = tensor<string, []>("input_473_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [2]> input_473_strides_0 = const()[name = tensor<string, []>("input_473_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> input_473_dilations_0 = const()[name = tensor<string, []>("input_473_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<fp16, [1024, 1, 1, 9]> const_302_to_fp16 = const()[name = tensor<string, []>("const_302_to_fp16"), val = tensor<fp16, [1024, 1, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(897356416)))];
             tensor<fp16, [1024]> const_303_to_fp16 = const()[name = tensor<string, []>("const_303_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(897374912)))];
-            tensor<fp16, [1, 1024, 1, 188]> input_473_cast_fp16 = conv(bias = const_303_to_fp16, dilations = input_471_dilations_0, groups = input_471_groups_0, pad = input_471_pad_0, pad_type = input_471_pad_type_0, strides = input_471_strides_0, weight = const_302_to_fp16, x = input_469_cast_fp16)[name = tensor<string, []>("input_473_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_475_cast_fp16 = silu(x = input_473_cast_fp16)[name = tensor<string, []>("input_475_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_475_cast_fp16 = conv(bias = const_303_to_fp16, dilations = input_473_dilations_0, groups = input_473_groups_0, pad = input_473_pad_0, pad_type = input_473_pad_type_0, strides = input_473_strides_0, weight = const_302_to_fp16, x = input_471_cast_fp16)[name = tensor<string, []>("input_475_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_477_cast_fp16 = silu(x = input_475_cast_fp16)[name = tensor<string, []>("input_477_cast_fp16")];
             tensor<string, []> x_107_pad_type_0 = const()[name = tensor<string, []>("x_107_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_107_strides_0 = const()[name = tensor<string, []>("x_107_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_107_pad_0 = const()[name = tensor<string, []>("x_107_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_107_dilations_0 = const()[name = tensor<string, []>("x_107_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_107_groups_0 = const()[name = tensor<string, []>("x_107_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_17_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("layers_17_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(897377024)))];
-            tensor<fp16, [1, 1024, 1, 188]> x_107_cast_fp16 = conv(dilations = x_107_dilations_0, groups = x_107_groups_0, pad = x_107_pad_0, pad_type = x_107_pad_type_0, strides = x_107_strides_0, weight = layers_17_conv_pointwise_conv2_weight_to_fp16, x = input_475_cast_fp16)[name = tensor<string, []>("x_107_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> x_107_cast_fp16 = conv(dilations = x_107_dilations_0, groups = x_107_groups_0, pad = x_107_pad_0, pad_type = x_107_pad_type_0, strides = x_107_strides_0, weight = layers_17_conv_pointwise_conv2_weight_to_fp16, x = input_477_cast_fp16)[name = tensor<string, []>("x_107_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_177_cast_fp16 = add(x = inputs_175_cast_fp16, y = x_107_cast_fp16)[name = tensor<string, []>("inputs_177_cast_fp16")];
             tensor<int32, [1]> out_177_axes_0 = const()[name = tensor<string, []>("out_177_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_5157_to_fp16 = const()[name = tensor<string, []>("op_5157_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_177_cast_fp16 = layer_norm(axes = out_177_axes_0, epsilon = var_5157_to_fp16, x = inputs_177_cast_fp16)[name = tensor<string, []>("out_177_cast_fp16")];
-            tensor<fp16, [1024]> input_477_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_477_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(899474240)))];
-            tensor<fp16, [1024]> input_477_beta_0_to_fp16 = const()[name = tensor<string, []>("input_477_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(899476352)))];
-            tensor<fp16, []> input_477_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_477_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_477_cast_fp16 = batch_norm(beta = input_477_beta_0_to_fp16, epsilon = input_477_epsilon_0_to_fp16, gamma = input_477_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_177_cast_fp16)[name = tensor<string, []>("input_477_cast_fp16")];
-            tensor<string, []> input_479_pad_type_0 = const()[name = tensor<string, []>("input_479_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_479_strides_0 = const()[name = tensor<string, []>("input_479_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_479_pad_0 = const()[name = tensor<string, []>("input_479_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_479_dilations_0 = const()[name = tensor<string, []>("input_479_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_479_groups_0 = const()[name = tensor<string, []>("input_479_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_5155_to_fp16 = const()[name = tensor<string, []>("op_5155_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_177_cast_fp16 = layer_norm(axes = out_177_axes_0, epsilon = var_5155_to_fp16, x = inputs_177_cast_fp16)[name = tensor<string, []>("out_177_cast_fp16")];
+            tensor<fp16, [1024]> input_479_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_479_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(899474240)))];
+            tensor<fp16, [1024]> input_479_beta_0_to_fp16 = const()[name = tensor<string, []>("input_479_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(899476352)))];
+            tensor<fp16, []> input_479_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_479_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_479_cast_fp16 = batch_norm(beta = input_479_beta_0_to_fp16, epsilon = input_479_epsilon_0_to_fp16, gamma = input_479_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_177_cast_fp16)[name = tensor<string, []>("input_479_cast_fp16")];
+            tensor<string, []> input_481_pad_type_0 = const()[name = tensor<string, []>("input_481_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_481_strides_0 = const()[name = tensor<string, []>("input_481_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_481_pad_0 = const()[name = tensor<string, []>("input_481_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_481_dilations_0 = const()[name = tensor<string, []>("input_481_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_481_groups_0 = const()[name = tensor<string, []>("input_481_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_17_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_17_feed_forward2_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(899478464)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_479_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_479_dilations_0, groups = input_479_groups_0, pad = input_479_pad_0, pad_type = input_479_pad_type_0, strides = input_479_strides_0, weight = layers_17_feed_forward2_fc1_weight_to_fp16, x = input_477_cast_fp16)[name = tensor<string, []>("input_479_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_481_cast_fp16 = silu(x = input_479_cast_fp16)[name = tensor<string, []>("input_481_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_481_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_481_dilations_0, groups = input_481_groups_0, pad = input_481_pad_0, pad_type = input_481_pad_type_0, strides = input_481_strides_0, weight = layers_17_feed_forward2_fc1_weight_to_fp16, x = input_479_cast_fp16)[name = tensor<string, []>("input_481_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_483_cast_fp16 = silu(x = input_481_cast_fp16)[name = tensor<string, []>("input_483_cast_fp16")];
             tensor<string, []> x_109_pad_type_0 = const()[name = tensor<string, []>("x_109_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_109_strides_0 = const()[name = tensor<string, []>("x_109_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_109_pad_0 = const()[name = tensor<string, []>("x_109_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_109_dilations_0 = const()[name = tensor<string, []>("x_109_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_109_groups_0 = const()[name = tensor<string, []>("x_109_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_5185_weight_0_to_fp16 = const()[name = tensor<string, []>("op_5185_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(907867136)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_5185_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_109_dilations_0, groups = x_109_groups_0, pad = x_109_pad_0, pad_type = x_109_pad_type_0, strides = x_109_strides_0, weight = var_5185_weight_0_to_fp16, x = input_481_cast_fp16)[name = tensor<string, []>("op_5185_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_179_cast_fp16 = add(x = inputs_177_cast_fp16, y = var_5185_cast_fp16)[name = tensor<string, []>("inputs_179_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_5183_weight_0_to_fp16 = const()[name = tensor<string, []>("op_5183_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(907867136)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_5183_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_109_dilations_0, groups = x_109_groups_0, pad = x_109_pad_0, pad_type = x_109_pad_type_0, strides = x_109_strides_0, weight = var_5183_weight_0_to_fp16, x = input_483_cast_fp16)[name = tensor<string, []>("op_5183_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_179_cast_fp16 = add(x = inputs_177_cast_fp16, y = var_5183_cast_fp16)[name = tensor<string, []>("inputs_179_cast_fp16")];
             tensor<int32, [1]> out_179_axes_0 = const()[name = tensor<string, []>("out_179_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_5195_to_fp16 = const()[name = tensor<string, []>("op_5195_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_179_cast_fp16 = layer_norm(axes = out_179_axes_0, epsilon = var_5195_to_fp16, x = inputs_179_cast_fp16)[name = tensor<string, []>("out_179_cast_fp16")];
+            tensor<fp16, []> var_5193_to_fp16 = const()[name = tensor<string, []>("op_5193_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_179_cast_fp16 = layer_norm(axes = out_179_axes_0, epsilon = var_5193_to_fp16, x = inputs_179_cast_fp16)[name = tensor<string, []>("out_179_cast_fp16")];
             tensor<fp16, [1024]> inputs_181_gamma_0_to_fp16 = const()[name = tensor<string, []>("inputs_181_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(916255808)))];
             tensor<fp16, [1024]> inputs_181_beta_0_to_fp16 = const()[name = tensor<string, []>("inputs_181_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(916257920)))];
             tensor<fp16, []> inputs_181_epsilon_0_to_fp16 = const()[name = tensor<string, []>("inputs_181_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_181_cast_fp16 = batch_norm(beta = inputs_181_beta_0_to_fp16, epsilon = inputs_181_epsilon_0_to_fp16, gamma = inputs_181_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_179_cast_fp16)[name = tensor<string, []>("inputs_181_cast_fp16")];
-            tensor<int32, []> var_5209 = const()[name = tensor<string, []>("op_5209"), val = tensor<int32, []>(3)];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_181_cast_fp16 = batch_norm(beta = inputs_181_beta_0_to_fp16, epsilon = inputs_181_epsilon_0_to_fp16, gamma = inputs_181_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_179_cast_fp16)[name = tensor<string, []>("inputs_181_cast_fp16")];
+            tensor<int32, []> var_5207 = const()[name = tensor<string, []>("op_5207"), val = tensor<int32, []>(3)];
             tensor<int32, [1]> out_181_axes_0 = const()[name = tensor<string, []>("out_181_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_5240_to_fp16 = const()[name = tensor<string, []>("op_5240_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_181_cast_fp16 = layer_norm(axes = out_181_axes_0, epsilon = var_5240_to_fp16, x = inputs_181_cast_fp16)[name = tensor<string, []>("out_181_cast_fp16")];
-            tensor<fp16, [1024]> input_483_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_483_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(916260032)))];
-            tensor<fp16, [1024]> input_483_beta_0_to_fp16 = const()[name = tensor<string, []>("input_483_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(916262144)))];
-            tensor<fp16, []> input_483_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_483_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_483_cast_fp16 = batch_norm(beta = input_483_beta_0_to_fp16, epsilon = input_483_epsilon_0_to_fp16, gamma = input_483_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_181_cast_fp16)[name = tensor<string, []>("input_483_cast_fp16")];
-            tensor<string, []> input_485_pad_type_0 = const()[name = tensor<string, []>("input_485_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_485_strides_0 = const()[name = tensor<string, []>("input_485_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_485_pad_0 = const()[name = tensor<string, []>("input_485_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_485_dilations_0 = const()[name = tensor<string, []>("input_485_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_485_groups_0 = const()[name = tensor<string, []>("input_485_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_5238_to_fp16 = const()[name = tensor<string, []>("op_5238_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_181_cast_fp16 = layer_norm(axes = out_181_axes_0, epsilon = var_5238_to_fp16, x = inputs_181_cast_fp16)[name = tensor<string, []>("out_181_cast_fp16")];
+            tensor<fp16, [1024]> input_485_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_485_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(916260032)))];
+            tensor<fp16, [1024]> input_485_beta_0_to_fp16 = const()[name = tensor<string, []>("input_485_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(916262144)))];
+            tensor<fp16, []> input_485_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_485_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_485_cast_fp16 = batch_norm(beta = input_485_beta_0_to_fp16, epsilon = input_485_epsilon_0_to_fp16, gamma = input_485_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_181_cast_fp16)[name = tensor<string, []>("input_485_cast_fp16")];
+            tensor<string, []> input_487_pad_type_0 = const()[name = tensor<string, []>("input_487_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_487_strides_0 = const()[name = tensor<string, []>("input_487_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_487_pad_0 = const()[name = tensor<string, []>("input_487_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_487_dilations_0 = const()[name = tensor<string, []>("input_487_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_487_groups_0 = const()[name = tensor<string, []>("input_487_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_18_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_18_feed_forward1_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(916264256)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_485_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_485_dilations_0, groups = input_485_groups_0, pad = input_485_pad_0, pad_type = input_485_pad_type_0, strides = input_485_strides_0, weight = layers_18_feed_forward1_fc1_weight_to_fp16, x = input_483_cast_fp16)[name = tensor<string, []>("input_485_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_487_cast_fp16 = silu(x = input_485_cast_fp16)[name = tensor<string, []>("input_487_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_487_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_487_dilations_0, groups = input_487_groups_0, pad = input_487_pad_0, pad_type = input_487_pad_type_0, strides = input_487_strides_0, weight = layers_18_feed_forward1_fc1_weight_to_fp16, x = input_485_cast_fp16)[name = tensor<string, []>("input_487_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_489_cast_fp16 = silu(x = input_487_cast_fp16)[name = tensor<string, []>("input_489_cast_fp16")];
             tensor<string, []> x_111_pad_type_0 = const()[name = tensor<string, []>("x_111_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_111_strides_0 = const()[name = tensor<string, []>("x_111_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_111_pad_0 = const()[name = tensor<string, []>("x_111_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_111_dilations_0 = const()[name = tensor<string, []>("x_111_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_111_groups_0 = const()[name = tensor<string, []>("x_111_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_5268_weight_0_to_fp16 = const()[name = tensor<string, []>("op_5268_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(924652928)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_5268_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_111_dilations_0, groups = x_111_groups_0, pad = x_111_pad_0, pad_type = x_111_pad_type_0, strides = x_111_strides_0, weight = var_5268_weight_0_to_fp16, x = input_487_cast_fp16)[name = tensor<string, []>("op_5268_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_183_cast_fp16 = add(x = inputs_181_cast_fp16, y = var_5268_cast_fp16)[name = tensor<string, []>("inputs_183_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_5266_weight_0_to_fp16 = const()[name = tensor<string, []>("op_5266_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(924652928)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_5266_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_111_dilations_0, groups = x_111_groups_0, pad = x_111_pad_0, pad_type = x_111_pad_type_0, strides = x_111_strides_0, weight = var_5266_weight_0_to_fp16, x = input_489_cast_fp16)[name = tensor<string, []>("op_5266_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_183_cast_fp16 = add(x = inputs_181_cast_fp16, y = var_5266_cast_fp16)[name = tensor<string, []>("inputs_183_cast_fp16")];
             tensor<int32, [1]> out_183_axes_0 = const()[name = tensor<string, []>("out_183_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_5278_to_fp16 = const()[name = tensor<string, []>("op_5278_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_183_cast_fp16 = layer_norm(axes = out_183_axes_0, epsilon = var_5278_to_fp16, x = inputs_183_cast_fp16)[name = tensor<string, []>("out_183_cast_fp16")];
+            tensor<fp16, []> var_5276_to_fp16 = const()[name = tensor<string, []>("op_5276_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_183_cast_fp16 = layer_norm(axes = out_183_axes_0, epsilon = var_5276_to_fp16, x = inputs_183_cast_fp16)[name = tensor<string, []>("out_183_cast_fp16")];
             tensor<fp16, [1024]> obj_75_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_75_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(933041600)))];
             tensor<fp16, [1024]> obj_75_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_75_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(933043712)))];
             tensor<fp16, []> obj_75_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_75_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> obj_75_cast_fp16 = batch_norm(beta = obj_75_beta_0_to_fp16, epsilon = obj_75_epsilon_0_to_fp16, gamma = obj_75_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_183_cast_fp16)[name = tensor<string, []>("obj_75_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_75_cast_fp16 = batch_norm(beta = obj_75_beta_0_to_fp16, epsilon = obj_75_epsilon_0_to_fp16, gamma = obj_75_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_183_cast_fp16)[name = tensor<string, []>("obj_75_cast_fp16")];
             tensor<string, []> query_73_pad_type_0 = const()[name = tensor<string, []>("query_73_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> query_73_strides_0 = const()[name = tensor<string, []>("query_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> query_73_pad_0 = const()[name = tensor<string, []>("query_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> query_73_dilations_0 = const()[name = tensor<string, []>("query_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> query_73_groups_0 = const()[name = tensor<string, []>("query_73_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_18_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_18_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(933045824)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_73_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = query_73_dilations_0, groups = query_73_groups_0, pad = query_73_pad_0, pad_type = query_73_pad_type_0, strides = query_73_strides_0, weight = layers_18_self_attn_q_proj_weight_to_fp16, x = obj_75_cast_fp16)[name = tensor<string, []>("query_73_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> query_73_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_73_dilations_0, groups = query_73_groups_0, pad = query_73_pad_0, pad_type = query_73_pad_type_0, strides = query_73_strides_0, weight = layers_18_self_attn_q_proj_weight_to_fp16, x = obj_75_cast_fp16)[name = tensor<string, []>("query_73_cast_fp16")];
             tensor<string, []> key_37_pad_type_0 = const()[name = tensor<string, []>("key_37_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> key_37_strides_0 = const()[name = tensor<string, []>("key_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> key_37_pad_0 = const()[name = tensor<string, []>("key_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
@@ -3352,173 +3354,173 @@ program(1.0)
             tensor<int32, [2]> value_37_dilations_0 = const()[name = tensor<string, []>("value_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> value_37_groups_0 = const()[name = tensor<string, []>("value_37_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_18_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_18_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(937240256)))];
-            tensor<fp16, [1, 1024, 1, 188]> value_37_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = value_37_dilations_0, groups = value_37_groups_0, pad = value_37_pad_0, pad_type = value_37_pad_type_0, strides = value_37_strides_0, weight = layers_18_self_attn_v_proj_weight_to_fp16, x = obj_75_cast_fp16)[name = tensor<string, []>("value_37_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_5316_to_fp16 = const()[name = tensor<string, []>("op_5316_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(939337472)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_75_cast_fp16 = add(x = query_73_cast_fp16, y = var_5316_to_fp16)[name = tensor<string, []>("query_75_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_5319_to_fp16 = const()[name = tensor<string, []>("op_5319_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(939339584)))];
-            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_37_cast_fp16 = add(x = query_73_cast_fp16, y = var_5319_to_fp16)[name = tensor<string, []>("q_with_bias_v_37_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> value_37_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_37_dilations_0, groups = value_37_groups_0, pad = value_37_pad_0, pad_type = value_37_pad_type_0, strides = value_37_strides_0, weight = layers_18_self_attn_v_proj_weight_to_fp16, x = obj_75_cast_fp16)[name = tensor<string, []>("value_37_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_5314_to_fp16 = const()[name = tensor<string, []>("op_5314_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(939337472)))];
+            tensor<fp16, [1, 1024, 1, 188]> query_75_cast_fp16 = add(x = query_73_cast_fp16, y = var_5314_to_fp16)[name = tensor<string, []>("query_75_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_5317_to_fp16 = const()[name = tensor<string, []>("op_5317_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(939339584)))];
+            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_37_cast_fp16 = add(x = query_73_cast_fp16, y = var_5317_to_fp16)[name = tensor<string, []>("q_with_bias_v_37_cast_fp16")];
             tensor<string, []> p_37_pad_type_0 = const()[name = tensor<string, []>("p_37_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> p_37_strides_0 = const()[name = tensor<string, []>("p_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> p_37_pad_0 = const()[name = tensor<string, []>("p_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> p_37_dilations_0 = const()[name = tensor<string, []>("p_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> p_37_groups_0 = const()[name = tensor<string, []>("p_37_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_18_self_attn_linear_pos_weight_to_fp16 = const()[name = tensor<string, []>("layers_18_self_attn_linear_pos_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(939341696)))];
-            tensor<fp16, [1, 1024, 1, 375]> p_37_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = p_37_dilations_0, groups = p_37_groups_0, pad = p_37_pad_0, pad_type = p_37_pad_type_0, strides = p_37_strides_0, weight = layers_18_self_attn_linear_pos_weight_to_fp16, x = pos_enc_to_fp16)[name = tensor<string, []>("p_37_cast_fp16")];
-            tensor<int32, [4]> var_5330 = const()[name = tensor<string, []>("op_5330"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_5331_cast_fp16 = reshape(shape = var_5330, x = q_with_bias_v_37_cast_fp16)[name = tensor<string, []>("op_5331_cast_fp16")];
-            tensor<int32, [4]> var_5332 = const()[name = tensor<string, []>("op_5332"), val = tensor<int32, [4]>([1, 8, 128, -1])];
-            tensor<fp16, [1, 8, 128, 375]> var_5333_cast_fp16 = reshape(shape = var_5332, x = p_37_cast_fp16)[name = tensor<string, []>("op_5333_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 375]> p_37_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_37_dilations_0, groups = p_37_groups_0, pad = p_37_pad_0, pad_type = p_37_pad_type_0, strides = p_37_strides_0, weight = layers_18_self_attn_linear_pos_weight_to_fp16, x = obj_3_cast_fp16)[name = tensor<string, []>("p_37_cast_fp16")];
+            tensor<int32, [4]> var_5328 = const()[name = tensor<string, []>("op_5328"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_5329_cast_fp16 = reshape(shape = var_5328, x = q_with_bias_v_37_cast_fp16)[name = tensor<string, []>("op_5329_cast_fp16")];
+            tensor<int32, [4]> var_5330 = const()[name = tensor<string, []>("op_5330"), val = tensor<int32, [4]>([1, 8, 128, -1])];
+            tensor<fp16, [1, 8, 128, 375]> var_5331_cast_fp16 = reshape(shape = var_5330, x = p_37_cast_fp16)[name = tensor<string, []>("op_5331_cast_fp16")];
             tensor<bool, []> matrix_bd_145_transpose_x_0 = const()[name = tensor<string, []>("matrix_bd_145_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> matrix_bd_145_transpose_y_0 = const()[name = tensor<string, []>("matrix_bd_145_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_145_cast_fp16 = matmul(transpose_x = matrix_bd_145_transpose_x_0, transpose_y = matrix_bd_145_transpose_y_0, x = var_5331_cast_fp16, y = var_5333_cast_fp16)[name = tensor<string, []>("matrix_bd_145_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_145_cast_fp16 = matmul(transpose_x = matrix_bd_145_transpose_x_0, transpose_y = matrix_bd_145_transpose_y_0, x = var_5329_cast_fp16, y = var_5331_cast_fp16)[name = tensor<string, []>("matrix_bd_145_cast_fp16")];
             tensor<int32, [8]> matrix_bd_147_pad_0 = const()[name = tensor<string, []>("matrix_bd_147_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
             tensor<string, []> matrix_bd_147_mode_0 = const()[name = tensor<string, []>("matrix_bd_147_mode_0"), val = tensor<string, []>("constant")];
             tensor<fp16, []> const_208_to_fp16 = const()[name = tensor<string, []>("const_208_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
             tensor<fp16, [1, 8, 188, 376]> matrix_bd_147_cast_fp16 = pad(constant_val = const_208_to_fp16, mode = matrix_bd_147_mode_0, pad = matrix_bd_147_pad_0, x = matrix_bd_145_cast_fp16)[name = tensor<string, []>("matrix_bd_147_cast_fp16")];
-            tensor<int32, [4]> var_5342 = const()[name = tensor<string, []>("op_5342"), val = tensor<int32, [4]>([1, 8, -1, 188])];
-            tensor<fp16, [1, 8, 376, 188]> matrix_bd_149_cast_fp16 = reshape(shape = var_5342, x = matrix_bd_147_cast_fp16)[name = tensor<string, []>("matrix_bd_149_cast_fp16")];
-            tensor<int32, [4]> var_5346_begin_0 = const()[name = tensor<string, []>("op_5346_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
-            tensor<int32, [4]> var_5346_end_0 = const()[name = tensor<string, []>("op_5346_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
-            tensor<bool, [4]> var_5346_end_mask_0 = const()[name = tensor<string, []>("op_5346_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
-            tensor<fp16, [1, 8, 375, 188]> var_5346_cast_fp16 = slice_by_index(begin = var_5346_begin_0, end = var_5346_end_0, end_mask = var_5346_end_mask_0, x = matrix_bd_149_cast_fp16)[name = tensor<string, []>("op_5346_cast_fp16")];
-            tensor<int32, [4]> var_5347 = const()[name = tensor<string, []>("op_5347"), val = tensor<int32, [4]>([1, 8, 188, 375])];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_151_cast_fp16 = reshape(shape = var_5347, x = var_5346_cast_fp16)[name = tensor<string, []>("matrix_bd_151_cast_fp16")];
-            tensor<int32, [4]> var_5352_begin_0 = const()[name = tensor<string, []>("op_5352_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [4]> var_5352_end_0 = const()[name = tensor<string, []>("op_5352_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
-            tensor<bool, [4]> var_5352_end_mask_0 = const()[name = tensor<string, []>("op_5352_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
-            tensor<fp16, [1, 8, 188, 188]> var_5352_cast_fp16 = slice_by_index(begin = var_5352_begin_0, end = var_5352_end_0, end_mask = var_5352_end_mask_0, x = matrix_bd_151_cast_fp16)[name = tensor<string, []>("op_5352_cast_fp16")];
-            tensor<fp16, []> var_5353_to_fp16 = const()[name = tensor<string, []>("op_5353_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 188, 188]> qk_mask_37_cast_fp16 = mul(x = var_5352_cast_fp16, y = var_5353_to_fp16)[name = tensor<string, []>("qk_mask_37_cast_fp16")];
-            tensor<int32, [4]> var_5357 = const()[name = tensor<string, []>("op_5357"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> mh_q_37_cast_fp16 = reshape(shape = var_5357, x = query_75_cast_fp16)[name = tensor<string, []>("mh_q_37_cast_fp16")];
-            tensor<fp16, []> var_5359_to_fp16 = const()[name = tensor<string, []>("op_5359_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 128, 188]> var_5360_cast_fp16 = mul(x = mh_q_37_cast_fp16, y = var_5359_to_fp16)[name = tensor<string, []>("op_5360_cast_fp16")];
-            tensor<int32, [4]> var_5363 = const()[name = tensor<string, []>("op_5363"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_5364_cast_fp16 = reshape(shape = var_5363, x = key_37_cast_fp16)[name = tensor<string, []>("op_5364_cast_fp16")];
+            tensor<int32, [4]> var_5340 = const()[name = tensor<string, []>("op_5340"), val = tensor<int32, [4]>([1, 8, -1, 188])];
+            tensor<fp16, [1, 8, 376, 188]> matrix_bd_149_cast_fp16 = reshape(shape = var_5340, x = matrix_bd_147_cast_fp16)[name = tensor<string, []>("matrix_bd_149_cast_fp16")];
+            tensor<int32, [4]> var_5344_begin_0 = const()[name = tensor<string, []>("op_5344_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_5344_end_0 = const()[name = tensor<string, []>("op_5344_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
+            tensor<bool, [4]> var_5344_end_mask_0 = const()[name = tensor<string, []>("op_5344_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 375, 188]> var_5344_cast_fp16 = slice_by_index(begin = var_5344_begin_0, end = var_5344_end_0, end_mask = var_5344_end_mask_0, x = matrix_bd_149_cast_fp16)[name = tensor<string, []>("op_5344_cast_fp16")];
+            tensor<int32, [4]> var_5345 = const()[name = tensor<string, []>("op_5345"), val = tensor<int32, [4]>([1, 8, 188, 375])];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_151_cast_fp16 = reshape(shape = var_5345, x = var_5344_cast_fp16)[name = tensor<string, []>("matrix_bd_151_cast_fp16")];
+            tensor<int32, [4]> var_5350_begin_0 = const()[name = tensor<string, []>("op_5350_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_5350_end_0 = const()[name = tensor<string, []>("op_5350_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
+            tensor<bool, [4]> var_5350_end_mask_0 = const()[name = tensor<string, []>("op_5350_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 188, 188]> var_5350_cast_fp16 = slice_by_index(begin = var_5350_begin_0, end = var_5350_end_0, end_mask = var_5350_end_mask_0, x = matrix_bd_151_cast_fp16)[name = tensor<string, []>("op_5350_cast_fp16")];
+            tensor<fp16, []> var_5351_to_fp16 = const()[name = tensor<string, []>("op_5351_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 188, 188]> qk_mask_37_cast_fp16 = mul(x = var_5350_cast_fp16, y = var_5351_to_fp16)[name = tensor<string, []>("qk_mask_37_cast_fp16")];
+            tensor<int32, [4]> var_5355 = const()[name = tensor<string, []>("op_5355"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> mh_q_37_cast_fp16 = reshape(shape = var_5355, x = query_75_cast_fp16)[name = tensor<string, []>("mh_q_37_cast_fp16")];
+            tensor<fp16, []> var_5357_to_fp16 = const()[name = tensor<string, []>("op_5357_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 128, 188]> var_5358_cast_fp16 = mul(x = mh_q_37_cast_fp16, y = var_5357_to_fp16)[name = tensor<string, []>("op_5358_cast_fp16")];
+            tensor<int32, [4]> var_5361 = const()[name = tensor<string, []>("op_5361"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_5362_cast_fp16 = reshape(shape = var_5361, x = key_37_cast_fp16)[name = tensor<string, []>("op_5362_cast_fp16")];
             tensor<bool, []> mh_w_73_transpose_x_0 = const()[name = tensor<string, []>("mh_w_73_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> mh_w_73_transpose_y_0 = const()[name = tensor<string, []>("mh_w_73_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 188]> mh_w_73_cast_fp16 = matmul(transpose_x = mh_w_73_transpose_x_0, transpose_y = mh_w_73_transpose_y_0, x = var_5360_cast_fp16, y = var_5364_cast_fp16)[name = tensor<string, []>("mh_w_73_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> mh_w_73_cast_fp16 = matmul(transpose_x = mh_w_73_transpose_x_0, transpose_y = mh_w_73_transpose_y_0, x = var_5358_cast_fp16, y = var_5362_cast_fp16)[name = tensor<string, []>("mh_w_73_cast_fp16")];
             tensor<fp16, [1, 8, 188, 188]> mh_w_75_cast_fp16 = add(x = mh_w_73_cast_fp16, y = qk_mask_37_cast_fp16)[name = tensor<string, []>("mh_w_75_cast_fp16")];
-            tensor<fp16, [1, 8, 188, 188]> var_5368_cast_fp16 = softmax(axis = var_5209, x = mh_w_75_cast_fp16)[name = tensor<string, []>("op_5368_cast_fp16")];
-            tensor<int32, [4]> var_5369 = const()[name = tensor<string, []>("op_5369"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_5370_cast_fp16 = reshape(shape = var_5369, x = value_37_cast_fp16)[name = tensor<string, []>("op_5370_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> var_5366_cast_fp16 = softmax(axis = var_5207, x = mh_w_75_cast_fp16)[name = tensor<string, []>("op_5366_cast_fp16")];
+            tensor<int32, [4]> var_5367 = const()[name = tensor<string, []>("op_5367"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_5368_cast_fp16 = reshape(shape = var_5367, x = value_37_cast_fp16)[name = tensor<string, []>("op_5368_cast_fp16")];
             tensor<bool, []> attn_37_transpose_x_0 = const()[name = tensor<string, []>("attn_37_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> attn_37_transpose_y_0 = const()[name = tensor<string, []>("attn_37_transpose_y_0"), val = tensor<bool, []>(true)];
-            tensor<fp16, [1, 8, 128, 188]> attn_37_cast_fp16 = matmul(transpose_x = attn_37_transpose_x_0, transpose_y = attn_37_transpose_y_0, x = var_5370_cast_fp16, y = var_5368_cast_fp16)[name = tensor<string, []>("attn_37_cast_fp16")];
-            tensor<int32, [4]> var_5373 = const()[name = tensor<string, []>("op_5373"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
-            tensor<fp16, [1, 1024, 1, 188]> input_489_cast_fp16 = reshape(shape = var_5373, x = attn_37_cast_fp16)[name = tensor<string, []>("input_489_cast_fp16")];
+            tensor<fp16, [1, 8, 128, 188]> attn_37_cast_fp16 = matmul(transpose_x = attn_37_transpose_x_0, transpose_y = attn_37_transpose_y_0, x = var_5368_cast_fp16, y = var_5366_cast_fp16)[name = tensor<string, []>("attn_37_cast_fp16")];
+            tensor<int32, [4]> var_5371 = const()[name = tensor<string, []>("op_5371"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
+            tensor<fp16, [1, 1024, 1, 188]> input_491_cast_fp16 = reshape(shape = var_5371, x = attn_37_cast_fp16)[name = tensor<string, []>("input_491_cast_fp16")];
             tensor<string, []> obj_77_pad_type_0 = const()[name = tensor<string, []>("obj_77_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> obj_77_strides_0 = const()[name = tensor<string, []>("obj_77_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> obj_77_pad_0 = const()[name = tensor<string, []>("obj_77_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> obj_77_dilations_0 = const()[name = tensor<string, []>("obj_77_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> obj_77_groups_0 = const()[name = tensor<string, []>("obj_77_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_18_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_18_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(941438912)))];
-            tensor<fp16, [1, 1024, 1, 188]> obj_77_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = obj_77_dilations_0, groups = obj_77_groups_0, pad = obj_77_pad_0, pad_type = obj_77_pad_type_0, strides = obj_77_strides_0, weight = layers_18_self_attn_o_proj_weight_to_fp16, x = input_489_cast_fp16)[name = tensor<string, []>("obj_77_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_77_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_77_dilations_0, groups = obj_77_groups_0, pad = obj_77_pad_0, pad_type = obj_77_pad_type_0, strides = obj_77_strides_0, weight = layers_18_self_attn_o_proj_weight_to_fp16, x = input_491_cast_fp16)[name = tensor<string, []>("obj_77_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_185_cast_fp16 = add(x = inputs_183_cast_fp16, y = obj_77_cast_fp16)[name = tensor<string, []>("inputs_185_cast_fp16")];
             tensor<int32, [1]> out_185_axes_0 = const()[name = tensor<string, []>("out_185_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_5391_to_fp16 = const()[name = tensor<string, []>("op_5391_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_185_cast_fp16 = layer_norm(axes = out_185_axes_0, epsilon = var_5391_to_fp16, x = inputs_185_cast_fp16)[name = tensor<string, []>("out_185_cast_fp16")];
-            tensor<fp16, [1024]> input_491_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_491_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(943536128)))];
-            tensor<fp16, [1024]> input_491_beta_0_to_fp16 = const()[name = tensor<string, []>("input_491_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(943538240)))];
-            tensor<fp16, []> input_491_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_491_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_491_cast_fp16 = batch_norm(beta = input_491_beta_0_to_fp16, epsilon = input_491_epsilon_0_to_fp16, gamma = input_491_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_185_cast_fp16)[name = tensor<string, []>("input_491_cast_fp16")];
-            tensor<string, []> input_493_pad_type_0 = const()[name = tensor<string, []>("input_493_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_493_strides_0 = const()[name = tensor<string, []>("input_493_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_493_pad_0 = const()[name = tensor<string, []>("input_493_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_493_dilations_0 = const()[name = tensor<string, []>("input_493_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_493_groups_0 = const()[name = tensor<string, []>("input_493_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_5389_to_fp16 = const()[name = tensor<string, []>("op_5389_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_185_cast_fp16 = layer_norm(axes = out_185_axes_0, epsilon = var_5389_to_fp16, x = inputs_185_cast_fp16)[name = tensor<string, []>("out_185_cast_fp16")];
+            tensor<fp16, [1024]> input_493_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_493_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(943536128)))];
+            tensor<fp16, [1024]> input_493_beta_0_to_fp16 = const()[name = tensor<string, []>("input_493_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(943538240)))];
+            tensor<fp16, []> input_493_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_493_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_493_cast_fp16 = batch_norm(beta = input_493_beta_0_to_fp16, epsilon = input_493_epsilon_0_to_fp16, gamma = input_493_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_185_cast_fp16)[name = tensor<string, []>("input_493_cast_fp16")];
+            tensor<string, []> input_495_pad_type_0 = const()[name = tensor<string, []>("input_495_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_495_strides_0 = const()[name = tensor<string, []>("input_495_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_495_pad_0 = const()[name = tensor<string, []>("input_495_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_495_dilations_0 = const()[name = tensor<string, []>("input_495_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_495_groups_0 = const()[name = tensor<string, []>("input_495_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [2048, 1024, 1, 1]> layers_18_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("layers_18_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(943540352)))];
-            tensor<fp16, [1, 2048, 1, 188]> input_493_cast_fp16 = conv(dilations = input_493_dilations_0, groups = input_493_groups_0, pad = input_493_pad_0, pad_type = input_493_pad_type_0, strides = input_493_strides_0, weight = layers_18_conv_pointwise_conv1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor<string, []>("input_493_cast_fp16")];
-            tensor<int32, []> input_495_split_num_splits_0 = const()[name = tensor<string, []>("input_495_split_num_splits_0"), val = tensor<int32, []>(2)];
-            tensor<int32, []> input_495_split_axis_0 = const()[name = tensor<string, []>("input_495_split_axis_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1, 1024, 1, 188]> input_495_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_495_split_cast_fp16_1 = split(axis = input_495_split_axis_0, num_splits = input_495_split_num_splits_0, x = input_493_cast_fp16)[name = tensor<string, []>("input_495_split_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_495_split_1_sigmoid_cast_fp16 = sigmoid(x = input_495_split_cast_fp16_1)[name = tensor<string, []>("input_495_split_1_sigmoid_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_495_cast_fp16 = mul(x = input_495_split_cast_fp16_0, y = input_495_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_495_cast_fp16")];
-            tensor<string, []> input_497_pad_type_0 = const()[name = tensor<string, []>("input_497_pad_type_0"), val = tensor<string, []>("custom")];
-            tensor<int32, [4]> input_497_pad_0 = const()[name = tensor<string, []>("input_497_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
-            tensor<int32, []> input_497_groups_0 = const()[name = tensor<string, []>("input_497_groups_0"), val = tensor<int32, []>(1024)];
-            tensor<int32, [2]> input_497_strides_0 = const()[name = tensor<string, []>("input_497_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> input_497_dilations_0 = const()[name = tensor<string, []>("input_497_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [1, 2048, 1, 188]> input_495_cast_fp16 = conv(dilations = input_495_dilations_0, groups = input_495_groups_0, pad = input_495_pad_0, pad_type = input_495_pad_type_0, strides = input_495_strides_0, weight = layers_18_conv_pointwise_conv1_weight_to_fp16, x = input_493_cast_fp16)[name = tensor<string, []>("input_495_cast_fp16")];
+            tensor<int32, []> input_497_split_num_splits_0 = const()[name = tensor<string, []>("input_497_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> input_497_split_axis_0 = const()[name = tensor<string, []>("input_497_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 1, 188]> input_497_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_497_split_cast_fp16_1 = split(axis = input_497_split_axis_0, num_splits = input_497_split_num_splits_0, x = input_495_cast_fp16)[name = tensor<string, []>("input_497_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_497_split_1_sigmoid_cast_fp16 = sigmoid(x = input_497_split_cast_fp16_1)[name = tensor<string, []>("input_497_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_497_cast_fp16 = mul(x = input_497_split_cast_fp16_0, y = input_497_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_497_cast_fp16")];
+            tensor<string, []> input_499_pad_type_0 = const()[name = tensor<string, []>("input_499_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_499_pad_0 = const()[name = tensor<string, []>("input_499_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
+            tensor<int32, []> input_499_groups_0 = const()[name = tensor<string, []>("input_499_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [2]> input_499_strides_0 = const()[name = tensor<string, []>("input_499_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> input_499_dilations_0 = const()[name = tensor<string, []>("input_499_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<fp16, [1024, 1, 1, 9]> const_304_to_fp16 = const()[name = tensor<string, []>("const_304_to_fp16"), val = tensor<fp16, [1024, 1, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(947734720)))];
             tensor<fp16, [1024]> const_305_to_fp16 = const()[name = tensor<string, []>("const_305_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(947753216)))];
-            tensor<fp16, [1, 1024, 1, 188]> input_499_cast_fp16 = conv(bias = const_305_to_fp16, dilations = input_497_dilations_0, groups = input_497_groups_0, pad = input_497_pad_0, pad_type = input_497_pad_type_0, strides = input_497_strides_0, weight = const_304_to_fp16, x = input_495_cast_fp16)[name = tensor<string, []>("input_499_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_501_cast_fp16 = silu(x = input_499_cast_fp16)[name = tensor<string, []>("input_501_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_501_cast_fp16 = conv(bias = const_305_to_fp16, dilations = input_499_dilations_0, groups = input_499_groups_0, pad = input_499_pad_0, pad_type = input_499_pad_type_0, strides = input_499_strides_0, weight = const_304_to_fp16, x = input_497_cast_fp16)[name = tensor<string, []>("input_501_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_503_cast_fp16 = silu(x = input_501_cast_fp16)[name = tensor<string, []>("input_503_cast_fp16")];
             tensor<string, []> x_113_pad_type_0 = const()[name = tensor<string, []>("x_113_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_113_strides_0 = const()[name = tensor<string, []>("x_113_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_113_pad_0 = const()[name = tensor<string, []>("x_113_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_113_dilations_0 = const()[name = tensor<string, []>("x_113_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_113_groups_0 = const()[name = tensor<string, []>("x_113_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_18_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("layers_18_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(947755328)))];
-            tensor<fp16, [1, 1024, 1, 188]> x_113_cast_fp16 = conv(dilations = x_113_dilations_0, groups = x_113_groups_0, pad = x_113_pad_0, pad_type = x_113_pad_type_0, strides = x_113_strides_0, weight = layers_18_conv_pointwise_conv2_weight_to_fp16, x = input_501_cast_fp16)[name = tensor<string, []>("x_113_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> x_113_cast_fp16 = conv(dilations = x_113_dilations_0, groups = x_113_groups_0, pad = x_113_pad_0, pad_type = x_113_pad_type_0, strides = x_113_strides_0, weight = layers_18_conv_pointwise_conv2_weight_to_fp16, x = input_503_cast_fp16)[name = tensor<string, []>("x_113_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_187_cast_fp16 = add(x = inputs_185_cast_fp16, y = x_113_cast_fp16)[name = tensor<string, []>("inputs_187_cast_fp16")];
             tensor<int32, [1]> out_187_axes_0 = const()[name = tensor<string, []>("out_187_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_5439_to_fp16 = const()[name = tensor<string, []>("op_5439_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_187_cast_fp16 = layer_norm(axes = out_187_axes_0, epsilon = var_5439_to_fp16, x = inputs_187_cast_fp16)[name = tensor<string, []>("out_187_cast_fp16")];
-            tensor<fp16, [1024]> input_503_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_503_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(949852544)))];
-            tensor<fp16, [1024]> input_503_beta_0_to_fp16 = const()[name = tensor<string, []>("input_503_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(949854656)))];
-            tensor<fp16, []> input_503_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_503_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_503_cast_fp16 = batch_norm(beta = input_503_beta_0_to_fp16, epsilon = input_503_epsilon_0_to_fp16, gamma = input_503_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_187_cast_fp16)[name = tensor<string, []>("input_503_cast_fp16")];
-            tensor<string, []> input_505_pad_type_0 = const()[name = tensor<string, []>("input_505_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_505_strides_0 = const()[name = tensor<string, []>("input_505_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_505_pad_0 = const()[name = tensor<string, []>("input_505_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_505_dilations_0 = const()[name = tensor<string, []>("input_505_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_505_groups_0 = const()[name = tensor<string, []>("input_505_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_5437_to_fp16 = const()[name = tensor<string, []>("op_5437_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_187_cast_fp16 = layer_norm(axes = out_187_axes_0, epsilon = var_5437_to_fp16, x = inputs_187_cast_fp16)[name = tensor<string, []>("out_187_cast_fp16")];
+            tensor<fp16, [1024]> input_505_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_505_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(949852544)))];
+            tensor<fp16, [1024]> input_505_beta_0_to_fp16 = const()[name = tensor<string, []>("input_505_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(949854656)))];
+            tensor<fp16, []> input_505_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_505_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_505_cast_fp16 = batch_norm(beta = input_505_beta_0_to_fp16, epsilon = input_505_epsilon_0_to_fp16, gamma = input_505_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_187_cast_fp16)[name = tensor<string, []>("input_505_cast_fp16")];
+            tensor<string, []> input_507_pad_type_0 = const()[name = tensor<string, []>("input_507_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_507_strides_0 = const()[name = tensor<string, []>("input_507_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_507_pad_0 = const()[name = tensor<string, []>("input_507_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_507_dilations_0 = const()[name = tensor<string, []>("input_507_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_507_groups_0 = const()[name = tensor<string, []>("input_507_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_18_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_18_feed_forward2_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(949856768)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_505_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_505_dilations_0, groups = input_505_groups_0, pad = input_505_pad_0, pad_type = input_505_pad_type_0, strides = input_505_strides_0, weight = layers_18_feed_forward2_fc1_weight_to_fp16, x = input_503_cast_fp16)[name = tensor<string, []>("input_505_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_507_cast_fp16 = silu(x = input_505_cast_fp16)[name = tensor<string, []>("input_507_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_507_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_507_dilations_0, groups = input_507_groups_0, pad = input_507_pad_0, pad_type = input_507_pad_type_0, strides = input_507_strides_0, weight = layers_18_feed_forward2_fc1_weight_to_fp16, x = input_505_cast_fp16)[name = tensor<string, []>("input_507_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_509_cast_fp16 = silu(x = input_507_cast_fp16)[name = tensor<string, []>("input_509_cast_fp16")];
             tensor<string, []> x_115_pad_type_0 = const()[name = tensor<string, []>("x_115_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_115_strides_0 = const()[name = tensor<string, []>("x_115_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_115_pad_0 = const()[name = tensor<string, []>("x_115_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_115_dilations_0 = const()[name = tensor<string, []>("x_115_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_115_groups_0 = const()[name = tensor<string, []>("x_115_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_5467_weight_0_to_fp16 = const()[name = tensor<string, []>("op_5467_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(958245440)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_5467_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_115_dilations_0, groups = x_115_groups_0, pad = x_115_pad_0, pad_type = x_115_pad_type_0, strides = x_115_strides_0, weight = var_5467_weight_0_to_fp16, x = input_507_cast_fp16)[name = tensor<string, []>("op_5467_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_189_cast_fp16 = add(x = inputs_187_cast_fp16, y = var_5467_cast_fp16)[name = tensor<string, []>("inputs_189_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_5465_weight_0_to_fp16 = const()[name = tensor<string, []>("op_5465_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(958245440)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_5465_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_115_dilations_0, groups = x_115_groups_0, pad = x_115_pad_0, pad_type = x_115_pad_type_0, strides = x_115_strides_0, weight = var_5465_weight_0_to_fp16, x = input_509_cast_fp16)[name = tensor<string, []>("op_5465_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_189_cast_fp16 = add(x = inputs_187_cast_fp16, y = var_5465_cast_fp16)[name = tensor<string, []>("inputs_189_cast_fp16")];
             tensor<int32, [1]> out_189_axes_0 = const()[name = tensor<string, []>("out_189_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_5477_to_fp16 = const()[name = tensor<string, []>("op_5477_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_189_cast_fp16 = layer_norm(axes = out_189_axes_0, epsilon = var_5477_to_fp16, x = inputs_189_cast_fp16)[name = tensor<string, []>("out_189_cast_fp16")];
+            tensor<fp16, []> var_5475_to_fp16 = const()[name = tensor<string, []>("op_5475_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_189_cast_fp16 = layer_norm(axes = out_189_axes_0, epsilon = var_5475_to_fp16, x = inputs_189_cast_fp16)[name = tensor<string, []>("out_189_cast_fp16")];
             tensor<fp16, [1024]> inputs_191_gamma_0_to_fp16 = const()[name = tensor<string, []>("inputs_191_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(966634112)))];
             tensor<fp16, [1024]> inputs_191_beta_0_to_fp16 = const()[name = tensor<string, []>("inputs_191_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(966636224)))];
             tensor<fp16, []> inputs_191_epsilon_0_to_fp16 = const()[name = tensor<string, []>("inputs_191_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_191_cast_fp16 = batch_norm(beta = inputs_191_beta_0_to_fp16, epsilon = inputs_191_epsilon_0_to_fp16, gamma = inputs_191_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_189_cast_fp16)[name = tensor<string, []>("inputs_191_cast_fp16")];
-            tensor<int32, []> var_5491 = const()[name = tensor<string, []>("op_5491"), val = tensor<int32, []>(3)];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_191_cast_fp16 = batch_norm(beta = inputs_191_beta_0_to_fp16, epsilon = inputs_191_epsilon_0_to_fp16, gamma = inputs_191_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_189_cast_fp16)[name = tensor<string, []>("inputs_191_cast_fp16")];
+            tensor<int32, []> var_5489 = const()[name = tensor<string, []>("op_5489"), val = tensor<int32, []>(3)];
             tensor<int32, [1]> out_191_axes_0 = const()[name = tensor<string, []>("out_191_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_5522_to_fp16 = const()[name = tensor<string, []>("op_5522_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_191_cast_fp16 = layer_norm(axes = out_191_axes_0, epsilon = var_5522_to_fp16, x = inputs_191_cast_fp16)[name = tensor<string, []>("out_191_cast_fp16")];
-            tensor<fp16, [1024]> input_509_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_509_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(966638336)))];
-            tensor<fp16, [1024]> input_509_beta_0_to_fp16 = const()[name = tensor<string, []>("input_509_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(966640448)))];
-            tensor<fp16, []> input_509_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_509_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_509_cast_fp16 = batch_norm(beta = input_509_beta_0_to_fp16, epsilon = input_509_epsilon_0_to_fp16, gamma = input_509_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_191_cast_fp16)[name = tensor<string, []>("input_509_cast_fp16")];
-            tensor<string, []> input_511_pad_type_0 = const()[name = tensor<string, []>("input_511_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_511_strides_0 = const()[name = tensor<string, []>("input_511_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_511_pad_0 = const()[name = tensor<string, []>("input_511_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_511_dilations_0 = const()[name = tensor<string, []>("input_511_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_511_groups_0 = const()[name = tensor<string, []>("input_511_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_5520_to_fp16 = const()[name = tensor<string, []>("op_5520_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_191_cast_fp16 = layer_norm(axes = out_191_axes_0, epsilon = var_5520_to_fp16, x = inputs_191_cast_fp16)[name = tensor<string, []>("out_191_cast_fp16")];
+            tensor<fp16, [1024]> input_511_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_511_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(966638336)))];
+            tensor<fp16, [1024]> input_511_beta_0_to_fp16 = const()[name = tensor<string, []>("input_511_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(966640448)))];
+            tensor<fp16, []> input_511_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_511_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_511_cast_fp16 = batch_norm(beta = input_511_beta_0_to_fp16, epsilon = input_511_epsilon_0_to_fp16, gamma = input_511_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_191_cast_fp16)[name = tensor<string, []>("input_511_cast_fp16")];
+            tensor<string, []> input_513_pad_type_0 = const()[name = tensor<string, []>("input_513_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_513_strides_0 = const()[name = tensor<string, []>("input_513_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_513_pad_0 = const()[name = tensor<string, []>("input_513_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_513_dilations_0 = const()[name = tensor<string, []>("input_513_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_513_groups_0 = const()[name = tensor<string, []>("input_513_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_19_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_19_feed_forward1_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(966642560)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_511_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_511_dilations_0, groups = input_511_groups_0, pad = input_511_pad_0, pad_type = input_511_pad_type_0, strides = input_511_strides_0, weight = layers_19_feed_forward1_fc1_weight_to_fp16, x = input_509_cast_fp16)[name = tensor<string, []>("input_511_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_513_cast_fp16 = silu(x = input_511_cast_fp16)[name = tensor<string, []>("input_513_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_513_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_513_dilations_0, groups = input_513_groups_0, pad = input_513_pad_0, pad_type = input_513_pad_type_0, strides = input_513_strides_0, weight = layers_19_feed_forward1_fc1_weight_to_fp16, x = input_511_cast_fp16)[name = tensor<string, []>("input_513_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_515_cast_fp16 = silu(x = input_513_cast_fp16)[name = tensor<string, []>("input_515_cast_fp16")];
             tensor<string, []> x_117_pad_type_0 = const()[name = tensor<string, []>("x_117_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_117_strides_0 = const()[name = tensor<string, []>("x_117_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_117_pad_0 = const()[name = tensor<string, []>("x_117_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_117_dilations_0 = const()[name = tensor<string, []>("x_117_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_117_groups_0 = const()[name = tensor<string, []>("x_117_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_5550_weight_0_to_fp16 = const()[name = tensor<string, []>("op_5550_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(975031232)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_5550_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_117_dilations_0, groups = x_117_groups_0, pad = x_117_pad_0, pad_type = x_117_pad_type_0, strides = x_117_strides_0, weight = var_5550_weight_0_to_fp16, x = input_513_cast_fp16)[name = tensor<string, []>("op_5550_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_193_cast_fp16 = add(x = inputs_191_cast_fp16, y = var_5550_cast_fp16)[name = tensor<string, []>("inputs_193_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_5548_weight_0_to_fp16 = const()[name = tensor<string, []>("op_5548_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(975031232)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_5548_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_117_dilations_0, groups = x_117_groups_0, pad = x_117_pad_0, pad_type = x_117_pad_type_0, strides = x_117_strides_0, weight = var_5548_weight_0_to_fp16, x = input_515_cast_fp16)[name = tensor<string, []>("op_5548_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_193_cast_fp16 = add(x = inputs_191_cast_fp16, y = var_5548_cast_fp16)[name = tensor<string, []>("inputs_193_cast_fp16")];
             tensor<int32, [1]> out_193_axes_0 = const()[name = tensor<string, []>("out_193_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_5560_to_fp16 = const()[name = tensor<string, []>("op_5560_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_193_cast_fp16 = layer_norm(axes = out_193_axes_0, epsilon = var_5560_to_fp16, x = inputs_193_cast_fp16)[name = tensor<string, []>("out_193_cast_fp16")];
+            tensor<fp16, []> var_5558_to_fp16 = const()[name = tensor<string, []>("op_5558_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_193_cast_fp16 = layer_norm(axes = out_193_axes_0, epsilon = var_5558_to_fp16, x = inputs_193_cast_fp16)[name = tensor<string, []>("out_193_cast_fp16")];
             tensor<fp16, [1024]> obj_79_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_79_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(983419904)))];
             tensor<fp16, [1024]> obj_79_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_79_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(983422016)))];
             tensor<fp16, []> obj_79_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_79_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> obj_79_cast_fp16 = batch_norm(beta = obj_79_beta_0_to_fp16, epsilon = obj_79_epsilon_0_to_fp16, gamma = obj_79_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_193_cast_fp16)[name = tensor<string, []>("obj_79_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_79_cast_fp16 = batch_norm(beta = obj_79_beta_0_to_fp16, epsilon = obj_79_epsilon_0_to_fp16, gamma = obj_79_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_193_cast_fp16)[name = tensor<string, []>("obj_79_cast_fp16")];
             tensor<string, []> query_77_pad_type_0 = const()[name = tensor<string, []>("query_77_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> query_77_strides_0 = const()[name = tensor<string, []>("query_77_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> query_77_pad_0 = const()[name = tensor<string, []>("query_77_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> query_77_dilations_0 = const()[name = tensor<string, []>("query_77_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> query_77_groups_0 = const()[name = tensor<string, []>("query_77_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_19_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_19_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(983424128)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_77_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = query_77_dilations_0, groups = query_77_groups_0, pad = query_77_pad_0, pad_type = query_77_pad_type_0, strides = query_77_strides_0, weight = layers_19_self_attn_q_proj_weight_to_fp16, x = obj_79_cast_fp16)[name = tensor<string, []>("query_77_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> query_77_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_77_dilations_0, groups = query_77_groups_0, pad = query_77_pad_0, pad_type = query_77_pad_type_0, strides = query_77_strides_0, weight = layers_19_self_attn_q_proj_weight_to_fp16, x = obj_79_cast_fp16)[name = tensor<string, []>("query_77_cast_fp16")];
             tensor<string, []> key_39_pad_type_0 = const()[name = tensor<string, []>("key_39_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> key_39_strides_0 = const()[name = tensor<string, []>("key_39_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> key_39_pad_0 = const()[name = tensor<string, []>("key_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
@@ -3532,173 +3534,173 @@ program(1.0)
             tensor<int32, [2]> value_39_dilations_0 = const()[name = tensor<string, []>("value_39_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> value_39_groups_0 = const()[name = tensor<string, []>("value_39_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_19_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_19_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(987618560)))];
-            tensor<fp16, [1, 1024, 1, 188]> value_39_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = value_39_dilations_0, groups = value_39_groups_0, pad = value_39_pad_0, pad_type = value_39_pad_type_0, strides = value_39_strides_0, weight = layers_19_self_attn_v_proj_weight_to_fp16, x = obj_79_cast_fp16)[name = tensor<string, []>("value_39_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_5598_to_fp16 = const()[name = tensor<string, []>("op_5598_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(989715776)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_79_cast_fp16 = add(x = query_77_cast_fp16, y = var_5598_to_fp16)[name = tensor<string, []>("query_79_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_5601_to_fp16 = const()[name = tensor<string, []>("op_5601_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(989717888)))];
-            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_39_cast_fp16 = add(x = query_77_cast_fp16, y = var_5601_to_fp16)[name = tensor<string, []>("q_with_bias_v_39_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> value_39_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_39_dilations_0, groups = value_39_groups_0, pad = value_39_pad_0, pad_type = value_39_pad_type_0, strides = value_39_strides_0, weight = layers_19_self_attn_v_proj_weight_to_fp16, x = obj_79_cast_fp16)[name = tensor<string, []>("value_39_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_5596_to_fp16 = const()[name = tensor<string, []>("op_5596_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(989715776)))];
+            tensor<fp16, [1, 1024, 1, 188]> query_79_cast_fp16 = add(x = query_77_cast_fp16, y = var_5596_to_fp16)[name = tensor<string, []>("query_79_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_5599_to_fp16 = const()[name = tensor<string, []>("op_5599_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(989717888)))];
+            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_39_cast_fp16 = add(x = query_77_cast_fp16, y = var_5599_to_fp16)[name = tensor<string, []>("q_with_bias_v_39_cast_fp16")];
             tensor<string, []> p_39_pad_type_0 = const()[name = tensor<string, []>("p_39_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> p_39_strides_0 = const()[name = tensor<string, []>("p_39_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> p_39_pad_0 = const()[name = tensor<string, []>("p_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> p_39_dilations_0 = const()[name = tensor<string, []>("p_39_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> p_39_groups_0 = const()[name = tensor<string, []>("p_39_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_19_self_attn_linear_pos_weight_to_fp16 = const()[name = tensor<string, []>("layers_19_self_attn_linear_pos_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(989720000)))];
-            tensor<fp16, [1, 1024, 1, 375]> p_39_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = p_39_dilations_0, groups = p_39_groups_0, pad = p_39_pad_0, pad_type = p_39_pad_type_0, strides = p_39_strides_0, weight = layers_19_self_attn_linear_pos_weight_to_fp16, x = pos_enc_to_fp16)[name = tensor<string, []>("p_39_cast_fp16")];
-            tensor<int32, [4]> var_5612 = const()[name = tensor<string, []>("op_5612"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_5613_cast_fp16 = reshape(shape = var_5612, x = q_with_bias_v_39_cast_fp16)[name = tensor<string, []>("op_5613_cast_fp16")];
-            tensor<int32, [4]> var_5614 = const()[name = tensor<string, []>("op_5614"), val = tensor<int32, [4]>([1, 8, 128, -1])];
-            tensor<fp16, [1, 8, 128, 375]> var_5615_cast_fp16 = reshape(shape = var_5614, x = p_39_cast_fp16)[name = tensor<string, []>("op_5615_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 375]> p_39_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_39_dilations_0, groups = p_39_groups_0, pad = p_39_pad_0, pad_type = p_39_pad_type_0, strides = p_39_strides_0, weight = layers_19_self_attn_linear_pos_weight_to_fp16, x = obj_3_cast_fp16)[name = tensor<string, []>("p_39_cast_fp16")];
+            tensor<int32, [4]> var_5610 = const()[name = tensor<string, []>("op_5610"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_5611_cast_fp16 = reshape(shape = var_5610, x = q_with_bias_v_39_cast_fp16)[name = tensor<string, []>("op_5611_cast_fp16")];
+            tensor<int32, [4]> var_5612 = const()[name = tensor<string, []>("op_5612"), val = tensor<int32, [4]>([1, 8, 128, -1])];
+            tensor<fp16, [1, 8, 128, 375]> var_5613_cast_fp16 = reshape(shape = var_5612, x = p_39_cast_fp16)[name = tensor<string, []>("op_5613_cast_fp16")];
             tensor<bool, []> matrix_bd_153_transpose_x_0 = const()[name = tensor<string, []>("matrix_bd_153_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> matrix_bd_153_transpose_y_0 = const()[name = tensor<string, []>("matrix_bd_153_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_153_cast_fp16 = matmul(transpose_x = matrix_bd_153_transpose_x_0, transpose_y = matrix_bd_153_transpose_y_0, x = var_5613_cast_fp16, y = var_5615_cast_fp16)[name = tensor<string, []>("matrix_bd_153_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_153_cast_fp16 = matmul(transpose_x = matrix_bd_153_transpose_x_0, transpose_y = matrix_bd_153_transpose_y_0, x = var_5611_cast_fp16, y = var_5613_cast_fp16)[name = tensor<string, []>("matrix_bd_153_cast_fp16")];
             tensor<int32, [8]> matrix_bd_155_pad_0 = const()[name = tensor<string, []>("matrix_bd_155_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
             tensor<string, []> matrix_bd_155_mode_0 = const()[name = tensor<string, []>("matrix_bd_155_mode_0"), val = tensor<string, []>("constant")];
             tensor<fp16, []> const_219_to_fp16 = const()[name = tensor<string, []>("const_219_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
             tensor<fp16, [1, 8, 188, 376]> matrix_bd_155_cast_fp16 = pad(constant_val = const_219_to_fp16, mode = matrix_bd_155_mode_0, pad = matrix_bd_155_pad_0, x = matrix_bd_153_cast_fp16)[name = tensor<string, []>("matrix_bd_155_cast_fp16")];
-            tensor<int32, [4]> var_5624 = const()[name = tensor<string, []>("op_5624"), val = tensor<int32, [4]>([1, 8, -1, 188])];
-            tensor<fp16, [1, 8, 376, 188]> matrix_bd_157_cast_fp16 = reshape(shape = var_5624, x = matrix_bd_155_cast_fp16)[name = tensor<string, []>("matrix_bd_157_cast_fp16")];
-            tensor<int32, [4]> var_5628_begin_0 = const()[name = tensor<string, []>("op_5628_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
-            tensor<int32, [4]> var_5628_end_0 = const()[name = tensor<string, []>("op_5628_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
-            tensor<bool, [4]> var_5628_end_mask_0 = const()[name = tensor<string, []>("op_5628_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
-            tensor<fp16, [1, 8, 375, 188]> var_5628_cast_fp16 = slice_by_index(begin = var_5628_begin_0, end = var_5628_end_0, end_mask = var_5628_end_mask_0, x = matrix_bd_157_cast_fp16)[name = tensor<string, []>("op_5628_cast_fp16")];
-            tensor<int32, [4]> var_5629 = const()[name = tensor<string, []>("op_5629"), val = tensor<int32, [4]>([1, 8, 188, 375])];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_159_cast_fp16 = reshape(shape = var_5629, x = var_5628_cast_fp16)[name = tensor<string, []>("matrix_bd_159_cast_fp16")];
-            tensor<int32, [4]> var_5634_begin_0 = const()[name = tensor<string, []>("op_5634_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [4]> var_5634_end_0 = const()[name = tensor<string, []>("op_5634_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
-            tensor<bool, [4]> var_5634_end_mask_0 = const()[name = tensor<string, []>("op_5634_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
-            tensor<fp16, [1, 8, 188, 188]> var_5634_cast_fp16 = slice_by_index(begin = var_5634_begin_0, end = var_5634_end_0, end_mask = var_5634_end_mask_0, x = matrix_bd_159_cast_fp16)[name = tensor<string, []>("op_5634_cast_fp16")];
-            tensor<fp16, []> var_5635_to_fp16 = const()[name = tensor<string, []>("op_5635_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 188, 188]> qk_mask_39_cast_fp16 = mul(x = var_5634_cast_fp16, y = var_5635_to_fp16)[name = tensor<string, []>("qk_mask_39_cast_fp16")];
-            tensor<int32, [4]> var_5639 = const()[name = tensor<string, []>("op_5639"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> mh_q_39_cast_fp16 = reshape(shape = var_5639, x = query_79_cast_fp16)[name = tensor<string, []>("mh_q_39_cast_fp16")];
-            tensor<fp16, []> var_5641_to_fp16 = const()[name = tensor<string, []>("op_5641_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 128, 188]> var_5642_cast_fp16 = mul(x = mh_q_39_cast_fp16, y = var_5641_to_fp16)[name = tensor<string, []>("op_5642_cast_fp16")];
-            tensor<int32, [4]> var_5645 = const()[name = tensor<string, []>("op_5645"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_5646_cast_fp16 = reshape(shape = var_5645, x = key_39_cast_fp16)[name = tensor<string, []>("op_5646_cast_fp16")];
+            tensor<int32, [4]> var_5622 = const()[name = tensor<string, []>("op_5622"), val = tensor<int32, [4]>([1, 8, -1, 188])];
+            tensor<fp16, [1, 8, 376, 188]> matrix_bd_157_cast_fp16 = reshape(shape = var_5622, x = matrix_bd_155_cast_fp16)[name = tensor<string, []>("matrix_bd_157_cast_fp16")];
+            tensor<int32, [4]> var_5626_begin_0 = const()[name = tensor<string, []>("op_5626_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_5626_end_0 = const()[name = tensor<string, []>("op_5626_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
+            tensor<bool, [4]> var_5626_end_mask_0 = const()[name = tensor<string, []>("op_5626_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 375, 188]> var_5626_cast_fp16 = slice_by_index(begin = var_5626_begin_0, end = var_5626_end_0, end_mask = var_5626_end_mask_0, x = matrix_bd_157_cast_fp16)[name = tensor<string, []>("op_5626_cast_fp16")];
+            tensor<int32, [4]> var_5627 = const()[name = tensor<string, []>("op_5627"), val = tensor<int32, [4]>([1, 8, 188, 375])];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_159_cast_fp16 = reshape(shape = var_5627, x = var_5626_cast_fp16)[name = tensor<string, []>("matrix_bd_159_cast_fp16")];
+            tensor<int32, [4]> var_5632_begin_0 = const()[name = tensor<string, []>("op_5632_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_5632_end_0 = const()[name = tensor<string, []>("op_5632_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
+            tensor<bool, [4]> var_5632_end_mask_0 = const()[name = tensor<string, []>("op_5632_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 188, 188]> var_5632_cast_fp16 = slice_by_index(begin = var_5632_begin_0, end = var_5632_end_0, end_mask = var_5632_end_mask_0, x = matrix_bd_159_cast_fp16)[name = tensor<string, []>("op_5632_cast_fp16")];
+            tensor<fp16, []> var_5633_to_fp16 = const()[name = tensor<string, []>("op_5633_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 188, 188]> qk_mask_39_cast_fp16 = mul(x = var_5632_cast_fp16, y = var_5633_to_fp16)[name = tensor<string, []>("qk_mask_39_cast_fp16")];
+            tensor<int32, [4]> var_5637 = const()[name = tensor<string, []>("op_5637"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> mh_q_39_cast_fp16 = reshape(shape = var_5637, x = query_79_cast_fp16)[name = tensor<string, []>("mh_q_39_cast_fp16")];
+            tensor<fp16, []> var_5639_to_fp16 = const()[name = tensor<string, []>("op_5639_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 128, 188]> var_5640_cast_fp16 = mul(x = mh_q_39_cast_fp16, y = var_5639_to_fp16)[name = tensor<string, []>("op_5640_cast_fp16")];
+            tensor<int32, [4]> var_5643 = const()[name = tensor<string, []>("op_5643"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_5644_cast_fp16 = reshape(shape = var_5643, x = key_39_cast_fp16)[name = tensor<string, []>("op_5644_cast_fp16")];
             tensor<bool, []> mh_w_77_transpose_x_0 = const()[name = tensor<string, []>("mh_w_77_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> mh_w_77_transpose_y_0 = const()[name = tensor<string, []>("mh_w_77_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 188]> mh_w_77_cast_fp16 = matmul(transpose_x = mh_w_77_transpose_x_0, transpose_y = mh_w_77_transpose_y_0, x = var_5642_cast_fp16, y = var_5646_cast_fp16)[name = tensor<string, []>("mh_w_77_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> mh_w_77_cast_fp16 = matmul(transpose_x = mh_w_77_transpose_x_0, transpose_y = mh_w_77_transpose_y_0, x = var_5640_cast_fp16, y = var_5644_cast_fp16)[name = tensor<string, []>("mh_w_77_cast_fp16")];
             tensor<fp16, [1, 8, 188, 188]> mh_w_79_cast_fp16 = add(x = mh_w_77_cast_fp16, y = qk_mask_39_cast_fp16)[name = tensor<string, []>("mh_w_79_cast_fp16")];
-            tensor<fp16, [1, 8, 188, 188]> var_5650_cast_fp16 = softmax(axis = var_5491, x = mh_w_79_cast_fp16)[name = tensor<string, []>("op_5650_cast_fp16")];
-            tensor<int32, [4]> var_5651 = const()[name = tensor<string, []>("op_5651"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_5652_cast_fp16 = reshape(shape = var_5651, x = value_39_cast_fp16)[name = tensor<string, []>("op_5652_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> var_5648_cast_fp16 = softmax(axis = var_5489, x = mh_w_79_cast_fp16)[name = tensor<string, []>("op_5648_cast_fp16")];
+            tensor<int32, [4]> var_5649 = const()[name = tensor<string, []>("op_5649"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_5650_cast_fp16 = reshape(shape = var_5649, x = value_39_cast_fp16)[name = tensor<string, []>("op_5650_cast_fp16")];
             tensor<bool, []> attn_39_transpose_x_0 = const()[name = tensor<string, []>("attn_39_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> attn_39_transpose_y_0 = const()[name = tensor<string, []>("attn_39_transpose_y_0"), val = tensor<bool, []>(true)];
-            tensor<fp16, [1, 8, 128, 188]> attn_39_cast_fp16 = matmul(transpose_x = attn_39_transpose_x_0, transpose_y = attn_39_transpose_y_0, x = var_5652_cast_fp16, y = var_5650_cast_fp16)[name = tensor<string, []>("attn_39_cast_fp16")];
-            tensor<int32, [4]> var_5655 = const()[name = tensor<string, []>("op_5655"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
-            tensor<fp16, [1, 1024, 1, 188]> input_515_cast_fp16 = reshape(shape = var_5655, x = attn_39_cast_fp16)[name = tensor<string, []>("input_515_cast_fp16")];
+            tensor<fp16, [1, 8, 128, 188]> attn_39_cast_fp16 = matmul(transpose_x = attn_39_transpose_x_0, transpose_y = attn_39_transpose_y_0, x = var_5650_cast_fp16, y = var_5648_cast_fp16)[name = tensor<string, []>("attn_39_cast_fp16")];
+            tensor<int32, [4]> var_5653 = const()[name = tensor<string, []>("op_5653"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
+            tensor<fp16, [1, 1024, 1, 188]> input_517_cast_fp16 = reshape(shape = var_5653, x = attn_39_cast_fp16)[name = tensor<string, []>("input_517_cast_fp16")];
             tensor<string, []> obj_81_pad_type_0 = const()[name = tensor<string, []>("obj_81_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> obj_81_strides_0 = const()[name = tensor<string, []>("obj_81_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> obj_81_pad_0 = const()[name = tensor<string, []>("obj_81_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> obj_81_dilations_0 = const()[name = tensor<string, []>("obj_81_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> obj_81_groups_0 = const()[name = tensor<string, []>("obj_81_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_19_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_19_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(991817216)))];
-            tensor<fp16, [1, 1024, 1, 188]> obj_81_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = obj_81_dilations_0, groups = obj_81_groups_0, pad = obj_81_pad_0, pad_type = obj_81_pad_type_0, strides = obj_81_strides_0, weight = layers_19_self_attn_o_proj_weight_to_fp16, x = input_515_cast_fp16)[name = tensor<string, []>("obj_81_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_81_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_81_dilations_0, groups = obj_81_groups_0, pad = obj_81_pad_0, pad_type = obj_81_pad_type_0, strides = obj_81_strides_0, weight = layers_19_self_attn_o_proj_weight_to_fp16, x = input_517_cast_fp16)[name = tensor<string, []>("obj_81_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_195_cast_fp16 = add(x = inputs_193_cast_fp16, y = obj_81_cast_fp16)[name = tensor<string, []>("inputs_195_cast_fp16")];
             tensor<int32, [1]> out_195_axes_0 = const()[name = tensor<string, []>("out_195_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_5673_to_fp16 = const()[name = tensor<string, []>("op_5673_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_195_cast_fp16 = layer_norm(axes = out_195_axes_0, epsilon = var_5673_to_fp16, x = inputs_195_cast_fp16)[name = tensor<string, []>("out_195_cast_fp16")];
-            tensor<fp16, [1024]> input_517_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_517_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(993914432)))];
-            tensor<fp16, [1024]> input_517_beta_0_to_fp16 = const()[name = tensor<string, []>("input_517_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(993916544)))];
-            tensor<fp16, []> input_517_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_517_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_517_cast_fp16 = batch_norm(beta = input_517_beta_0_to_fp16, epsilon = input_517_epsilon_0_to_fp16, gamma = input_517_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_195_cast_fp16)[name = tensor<string, []>("input_517_cast_fp16")];
-            tensor<string, []> input_519_pad_type_0 = const()[name = tensor<string, []>("input_519_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_519_strides_0 = const()[name = tensor<string, []>("input_519_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_519_pad_0 = const()[name = tensor<string, []>("input_519_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_519_dilations_0 = const()[name = tensor<string, []>("input_519_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_519_groups_0 = const()[name = tensor<string, []>("input_519_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_5671_to_fp16 = const()[name = tensor<string, []>("op_5671_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_195_cast_fp16 = layer_norm(axes = out_195_axes_0, epsilon = var_5671_to_fp16, x = inputs_195_cast_fp16)[name = tensor<string, []>("out_195_cast_fp16")];
+            tensor<fp16, [1024]> input_519_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_519_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(993914432)))];
+            tensor<fp16, [1024]> input_519_beta_0_to_fp16 = const()[name = tensor<string, []>("input_519_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(993916544)))];
+            tensor<fp16, []> input_519_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_519_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_519_cast_fp16 = batch_norm(beta = input_519_beta_0_to_fp16, epsilon = input_519_epsilon_0_to_fp16, gamma = input_519_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_195_cast_fp16)[name = tensor<string, []>("input_519_cast_fp16")];
+            tensor<string, []> input_521_pad_type_0 = const()[name = tensor<string, []>("input_521_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_521_strides_0 = const()[name = tensor<string, []>("input_521_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_521_pad_0 = const()[name = tensor<string, []>("input_521_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_521_dilations_0 = const()[name = tensor<string, []>("input_521_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_521_groups_0 = const()[name = tensor<string, []>("input_521_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [2048, 1024, 1, 1]> layers_19_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("layers_19_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(993918656)))];
-            tensor<fp16, [1, 2048, 1, 188]> input_519_cast_fp16 = conv(dilations = input_519_dilations_0, groups = input_519_groups_0, pad = input_519_pad_0, pad_type = input_519_pad_type_0, strides = input_519_strides_0, weight = layers_19_conv_pointwise_conv1_weight_to_fp16, x = input_517_cast_fp16)[name = tensor<string, []>("input_519_cast_fp16")];
-            tensor<int32, []> input_521_split_num_splits_0 = const()[name = tensor<string, []>("input_521_split_num_splits_0"), val = tensor<int32, []>(2)];
-            tensor<int32, []> input_521_split_axis_0 = const()[name = tensor<string, []>("input_521_split_axis_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1, 1024, 1, 188]> input_521_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_521_split_cast_fp16_1 = split(axis = input_521_split_axis_0, num_splits = input_521_split_num_splits_0, x = input_519_cast_fp16)[name = tensor<string, []>("input_521_split_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_521_split_1_sigmoid_cast_fp16 = sigmoid(x = input_521_split_cast_fp16_1)[name = tensor<string, []>("input_521_split_1_sigmoid_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_521_cast_fp16 = mul(x = input_521_split_cast_fp16_0, y = input_521_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_521_cast_fp16")];
-            tensor<string, []> input_523_pad_type_0 = const()[name = tensor<string, []>("input_523_pad_type_0"), val = tensor<string, []>("custom")];
-            tensor<int32, [4]> input_523_pad_0 = const()[name = tensor<string, []>("input_523_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
-            tensor<int32, []> input_523_groups_0 = const()[name = tensor<string, []>("input_523_groups_0"), val = tensor<int32, []>(1024)];
-            tensor<int32, [2]> input_523_strides_0 = const()[name = tensor<string, []>("input_523_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> input_523_dilations_0 = const()[name = tensor<string, []>("input_523_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [1, 2048, 1, 188]> input_521_cast_fp16 = conv(dilations = input_521_dilations_0, groups = input_521_groups_0, pad = input_521_pad_0, pad_type = input_521_pad_type_0, strides = input_521_strides_0, weight = layers_19_conv_pointwise_conv1_weight_to_fp16, x = input_519_cast_fp16)[name = tensor<string, []>("input_521_cast_fp16")];
+            tensor<int32, []> input_523_split_num_splits_0 = const()[name = tensor<string, []>("input_523_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> input_523_split_axis_0 = const()[name = tensor<string, []>("input_523_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 1, 188]> input_523_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_523_split_cast_fp16_1 = split(axis = input_523_split_axis_0, num_splits = input_523_split_num_splits_0, x = input_521_cast_fp16)[name = tensor<string, []>("input_523_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_523_split_1_sigmoid_cast_fp16 = sigmoid(x = input_523_split_cast_fp16_1)[name = tensor<string, []>("input_523_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_523_cast_fp16 = mul(x = input_523_split_cast_fp16_0, y = input_523_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_523_cast_fp16")];
+            tensor<string, []> input_525_pad_type_0 = const()[name = tensor<string, []>("input_525_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_525_pad_0 = const()[name = tensor<string, []>("input_525_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
+            tensor<int32, []> input_525_groups_0 = const()[name = tensor<string, []>("input_525_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [2]> input_525_strides_0 = const()[name = tensor<string, []>("input_525_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> input_525_dilations_0 = const()[name = tensor<string, []>("input_525_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<fp16, [1024, 1, 1, 9]> const_306_to_fp16 = const()[name = tensor<string, []>("const_306_to_fp16"), val = tensor<fp16, [1024, 1, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(998113024)))];
             tensor<fp16, [1024]> const_307_to_fp16 = const()[name = tensor<string, []>("const_307_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(998131520)))];
-            tensor<fp16, [1, 1024, 1, 188]> input_525_cast_fp16 = conv(bias = const_307_to_fp16, dilations = input_523_dilations_0, groups = input_523_groups_0, pad = input_523_pad_0, pad_type = input_523_pad_type_0, strides = input_523_strides_0, weight = const_306_to_fp16, x = input_521_cast_fp16)[name = tensor<string, []>("input_525_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_527_cast_fp16 = silu(x = input_525_cast_fp16)[name = tensor<string, []>("input_527_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_527_cast_fp16 = conv(bias = const_307_to_fp16, dilations = input_525_dilations_0, groups = input_525_groups_0, pad = input_525_pad_0, pad_type = input_525_pad_type_0, strides = input_525_strides_0, weight = const_306_to_fp16, x = input_523_cast_fp16)[name = tensor<string, []>("input_527_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_529_cast_fp16 = silu(x = input_527_cast_fp16)[name = tensor<string, []>("input_529_cast_fp16")];
             tensor<string, []> x_119_pad_type_0 = const()[name = tensor<string, []>("x_119_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_119_strides_0 = const()[name = tensor<string, []>("x_119_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_119_pad_0 = const()[name = tensor<string, []>("x_119_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_119_dilations_0 = const()[name = tensor<string, []>("x_119_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_119_groups_0 = const()[name = tensor<string, []>("x_119_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_19_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("layers_19_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(998133632)))];
-            tensor<fp16, [1, 1024, 1, 188]> x_119_cast_fp16 = conv(dilations = x_119_dilations_0, groups = x_119_groups_0, pad = x_119_pad_0, pad_type = x_119_pad_type_0, strides = x_119_strides_0, weight = layers_19_conv_pointwise_conv2_weight_to_fp16, x = input_527_cast_fp16)[name = tensor<string, []>("x_119_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> x_119_cast_fp16 = conv(dilations = x_119_dilations_0, groups = x_119_groups_0, pad = x_119_pad_0, pad_type = x_119_pad_type_0, strides = x_119_strides_0, weight = layers_19_conv_pointwise_conv2_weight_to_fp16, x = input_529_cast_fp16)[name = tensor<string, []>("x_119_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_197_cast_fp16 = add(x = inputs_195_cast_fp16, y = x_119_cast_fp16)[name = tensor<string, []>("inputs_197_cast_fp16")];
             tensor<int32, [1]> out_197_axes_0 = const()[name = tensor<string, []>("out_197_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_5721_to_fp16 = const()[name = tensor<string, []>("op_5721_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_197_cast_fp16 = layer_norm(axes = out_197_axes_0, epsilon = var_5721_to_fp16, x = inputs_197_cast_fp16)[name = tensor<string, []>("out_197_cast_fp16")];
-            tensor<fp16, [1024]> input_529_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_529_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1000230848)))];
-            tensor<fp16, [1024]> input_529_beta_0_to_fp16 = const()[name = tensor<string, []>("input_529_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1000232960)))];
-            tensor<fp16, []> input_529_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_529_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_529_cast_fp16 = batch_norm(beta = input_529_beta_0_to_fp16, epsilon = input_529_epsilon_0_to_fp16, gamma = input_529_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_197_cast_fp16)[name = tensor<string, []>("input_529_cast_fp16")];
-            tensor<string, []> input_531_pad_type_0 = const()[name = tensor<string, []>("input_531_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_531_strides_0 = const()[name = tensor<string, []>("input_531_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_531_pad_0 = const()[name = tensor<string, []>("input_531_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_531_dilations_0 = const()[name = tensor<string, []>("input_531_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_531_groups_0 = const()[name = tensor<string, []>("input_531_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_5719_to_fp16 = const()[name = tensor<string, []>("op_5719_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_197_cast_fp16 = layer_norm(axes = out_197_axes_0, epsilon = var_5719_to_fp16, x = inputs_197_cast_fp16)[name = tensor<string, []>("out_197_cast_fp16")];
+            tensor<fp16, [1024]> input_531_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_531_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1000230848)))];
+            tensor<fp16, [1024]> input_531_beta_0_to_fp16 = const()[name = tensor<string, []>("input_531_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1000232960)))];
+            tensor<fp16, []> input_531_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_531_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_531_cast_fp16 = batch_norm(beta = input_531_beta_0_to_fp16, epsilon = input_531_epsilon_0_to_fp16, gamma = input_531_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_197_cast_fp16)[name = tensor<string, []>("input_531_cast_fp16")];
+            tensor<string, []> input_533_pad_type_0 = const()[name = tensor<string, []>("input_533_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_533_strides_0 = const()[name = tensor<string, []>("input_533_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_533_pad_0 = const()[name = tensor<string, []>("input_533_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_533_dilations_0 = const()[name = tensor<string, []>("input_533_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_533_groups_0 = const()[name = tensor<string, []>("input_533_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_19_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_19_feed_forward2_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1000235072)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_531_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_531_dilations_0, groups = input_531_groups_0, pad = input_531_pad_0, pad_type = input_531_pad_type_0, strides = input_531_strides_0, weight = layers_19_feed_forward2_fc1_weight_to_fp16, x = input_529_cast_fp16)[name = tensor<string, []>("input_531_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_533_cast_fp16 = silu(x = input_531_cast_fp16)[name = tensor<string, []>("input_533_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_533_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_533_dilations_0, groups = input_533_groups_0, pad = input_533_pad_0, pad_type = input_533_pad_type_0, strides = input_533_strides_0, weight = layers_19_feed_forward2_fc1_weight_to_fp16, x = input_531_cast_fp16)[name = tensor<string, []>("input_533_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_535_cast_fp16 = silu(x = input_533_cast_fp16)[name = tensor<string, []>("input_535_cast_fp16")];
             tensor<string, []> x_121_pad_type_0 = const()[name = tensor<string, []>("x_121_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_121_strides_0 = const()[name = tensor<string, []>("x_121_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_121_pad_0 = const()[name = tensor<string, []>("x_121_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_121_dilations_0 = const()[name = tensor<string, []>("x_121_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_121_groups_0 = const()[name = tensor<string, []>("x_121_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_5749_weight_0_to_fp16 = const()[name = tensor<string, []>("op_5749_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1008623744)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_5749_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_121_dilations_0, groups = x_121_groups_0, pad = x_121_pad_0, pad_type = x_121_pad_type_0, strides = x_121_strides_0, weight = var_5749_weight_0_to_fp16, x = input_533_cast_fp16)[name = tensor<string, []>("op_5749_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_199_cast_fp16 = add(x = inputs_197_cast_fp16, y = var_5749_cast_fp16)[name = tensor<string, []>("inputs_199_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_5747_weight_0_to_fp16 = const()[name = tensor<string, []>("op_5747_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1008623744)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_5747_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_121_dilations_0, groups = x_121_groups_0, pad = x_121_pad_0, pad_type = x_121_pad_type_0, strides = x_121_strides_0, weight = var_5747_weight_0_to_fp16, x = input_535_cast_fp16)[name = tensor<string, []>("op_5747_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_199_cast_fp16 = add(x = inputs_197_cast_fp16, y = var_5747_cast_fp16)[name = tensor<string, []>("inputs_199_cast_fp16")];
             tensor<int32, [1]> out_199_axes_0 = const()[name = tensor<string, []>("out_199_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_5759_to_fp16 = const()[name = tensor<string, []>("op_5759_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_199_cast_fp16 = layer_norm(axes = out_199_axes_0, epsilon = var_5759_to_fp16, x = inputs_199_cast_fp16)[name = tensor<string, []>("out_199_cast_fp16")];
+            tensor<fp16, []> var_5757_to_fp16 = const()[name = tensor<string, []>("op_5757_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_199_cast_fp16 = layer_norm(axes = out_199_axes_0, epsilon = var_5757_to_fp16, x = inputs_199_cast_fp16)[name = tensor<string, []>("out_199_cast_fp16")];
             tensor<fp16, [1024]> inputs_201_gamma_0_to_fp16 = const()[name = tensor<string, []>("inputs_201_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1017012416)))];
             tensor<fp16, [1024]> inputs_201_beta_0_to_fp16 = const()[name = tensor<string, []>("inputs_201_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1017014528)))];
             tensor<fp16, []> inputs_201_epsilon_0_to_fp16 = const()[name = tensor<string, []>("inputs_201_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_201_cast_fp16 = batch_norm(beta = inputs_201_beta_0_to_fp16, epsilon = inputs_201_epsilon_0_to_fp16, gamma = inputs_201_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_199_cast_fp16)[name = tensor<string, []>("inputs_201_cast_fp16")];
-            tensor<int32, []> var_5773 = const()[name = tensor<string, []>("op_5773"), val = tensor<int32, []>(3)];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_201_cast_fp16 = batch_norm(beta = inputs_201_beta_0_to_fp16, epsilon = inputs_201_epsilon_0_to_fp16, gamma = inputs_201_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_199_cast_fp16)[name = tensor<string, []>("inputs_201_cast_fp16")];
+            tensor<int32, []> var_5771 = const()[name = tensor<string, []>("op_5771"), val = tensor<int32, []>(3)];
             tensor<int32, [1]> out_201_axes_0 = const()[name = tensor<string, []>("out_201_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_5804_to_fp16 = const()[name = tensor<string, []>("op_5804_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_201_cast_fp16 = layer_norm(axes = out_201_axes_0, epsilon = var_5804_to_fp16, x = inputs_201_cast_fp16)[name = tensor<string, []>("out_201_cast_fp16")];
-            tensor<fp16, [1024]> input_535_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_535_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1017016640)))];
-            tensor<fp16, [1024]> input_535_beta_0_to_fp16 = const()[name = tensor<string, []>("input_535_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1017018752)))];
-            tensor<fp16, []> input_535_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_535_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_535_cast_fp16 = batch_norm(beta = input_535_beta_0_to_fp16, epsilon = input_535_epsilon_0_to_fp16, gamma = input_535_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_201_cast_fp16)[name = tensor<string, []>("input_535_cast_fp16")];
-            tensor<string, []> input_537_pad_type_0 = const()[name = tensor<string, []>("input_537_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_537_strides_0 = const()[name = tensor<string, []>("input_537_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_537_pad_0 = const()[name = tensor<string, []>("input_537_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_537_dilations_0 = const()[name = tensor<string, []>("input_537_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_537_groups_0 = const()[name = tensor<string, []>("input_537_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_5802_to_fp16 = const()[name = tensor<string, []>("op_5802_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_201_cast_fp16 = layer_norm(axes = out_201_axes_0, epsilon = var_5802_to_fp16, x = inputs_201_cast_fp16)[name = tensor<string, []>("out_201_cast_fp16")];
+            tensor<fp16, [1024]> input_537_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_537_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1017016640)))];
+            tensor<fp16, [1024]> input_537_beta_0_to_fp16 = const()[name = tensor<string, []>("input_537_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1017018752)))];
+            tensor<fp16, []> input_537_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_537_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_537_cast_fp16 = batch_norm(beta = input_537_beta_0_to_fp16, epsilon = input_537_epsilon_0_to_fp16, gamma = input_537_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_201_cast_fp16)[name = tensor<string, []>("input_537_cast_fp16")];
+            tensor<string, []> input_539_pad_type_0 = const()[name = tensor<string, []>("input_539_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_539_strides_0 = const()[name = tensor<string, []>("input_539_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_539_pad_0 = const()[name = tensor<string, []>("input_539_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_539_dilations_0 = const()[name = tensor<string, []>("input_539_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_539_groups_0 = const()[name = tensor<string, []>("input_539_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_20_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_20_feed_forward1_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1017020864)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_537_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_537_dilations_0, groups = input_537_groups_0, pad = input_537_pad_0, pad_type = input_537_pad_type_0, strides = input_537_strides_0, weight = layers_20_feed_forward1_fc1_weight_to_fp16, x = input_535_cast_fp16)[name = tensor<string, []>("input_537_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_539_cast_fp16 = silu(x = input_537_cast_fp16)[name = tensor<string, []>("input_539_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_539_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_539_dilations_0, groups = input_539_groups_0, pad = input_539_pad_0, pad_type = input_539_pad_type_0, strides = input_539_strides_0, weight = layers_20_feed_forward1_fc1_weight_to_fp16, x = input_537_cast_fp16)[name = tensor<string, []>("input_539_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_541_cast_fp16 = silu(x = input_539_cast_fp16)[name = tensor<string, []>("input_541_cast_fp16")];
             tensor<string, []> x_123_pad_type_0 = const()[name = tensor<string, []>("x_123_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_123_strides_0 = const()[name = tensor<string, []>("x_123_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_123_pad_0 = const()[name = tensor<string, []>("x_123_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_123_dilations_0 = const()[name = tensor<string, []>("x_123_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_123_groups_0 = const()[name = tensor<string, []>("x_123_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_5832_weight_0_to_fp16 = const()[name = tensor<string, []>("op_5832_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1025409536)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_5832_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_123_dilations_0, groups = x_123_groups_0, pad = x_123_pad_0, pad_type = x_123_pad_type_0, strides = x_123_strides_0, weight = var_5832_weight_0_to_fp16, x = input_539_cast_fp16)[name = tensor<string, []>("op_5832_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_203_cast_fp16 = add(x = inputs_201_cast_fp16, y = var_5832_cast_fp16)[name = tensor<string, []>("inputs_203_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_5830_weight_0_to_fp16 = const()[name = tensor<string, []>("op_5830_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1025409536)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_5830_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_123_dilations_0, groups = x_123_groups_0, pad = x_123_pad_0, pad_type = x_123_pad_type_0, strides = x_123_strides_0, weight = var_5830_weight_0_to_fp16, x = input_541_cast_fp16)[name = tensor<string, []>("op_5830_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_203_cast_fp16 = add(x = inputs_201_cast_fp16, y = var_5830_cast_fp16)[name = tensor<string, []>("inputs_203_cast_fp16")];
             tensor<int32, [1]> out_203_axes_0 = const()[name = tensor<string, []>("out_203_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_5842_to_fp16 = const()[name = tensor<string, []>("op_5842_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_203_cast_fp16 = layer_norm(axes = out_203_axes_0, epsilon = var_5842_to_fp16, x = inputs_203_cast_fp16)[name = tensor<string, []>("out_203_cast_fp16")];
+            tensor<fp16, []> var_5840_to_fp16 = const()[name = tensor<string, []>("op_5840_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_203_cast_fp16 = layer_norm(axes = out_203_axes_0, epsilon = var_5840_to_fp16, x = inputs_203_cast_fp16)[name = tensor<string, []>("out_203_cast_fp16")];
             tensor<fp16, [1024]> obj_83_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_83_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1033798208)))];
             tensor<fp16, [1024]> obj_83_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_83_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1033800320)))];
             tensor<fp16, []> obj_83_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_83_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> obj_83_cast_fp16 = batch_norm(beta = obj_83_beta_0_to_fp16, epsilon = obj_83_epsilon_0_to_fp16, gamma = obj_83_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_203_cast_fp16)[name = tensor<string, []>("obj_83_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_83_cast_fp16 = batch_norm(beta = obj_83_beta_0_to_fp16, epsilon = obj_83_epsilon_0_to_fp16, gamma = obj_83_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_203_cast_fp16)[name = tensor<string, []>("obj_83_cast_fp16")];
             tensor<string, []> query_81_pad_type_0 = const()[name = tensor<string, []>("query_81_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> query_81_strides_0 = const()[name = tensor<string, []>("query_81_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> query_81_pad_0 = const()[name = tensor<string, []>("query_81_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> query_81_dilations_0 = const()[name = tensor<string, []>("query_81_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> query_81_groups_0 = const()[name = tensor<string, []>("query_81_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_20_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_20_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1033802432)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_81_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = query_81_dilations_0, groups = query_81_groups_0, pad = query_81_pad_0, pad_type = query_81_pad_type_0, strides = query_81_strides_0, weight = layers_20_self_attn_q_proj_weight_to_fp16, x = obj_83_cast_fp16)[name = tensor<string, []>("query_81_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> query_81_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_81_dilations_0, groups = query_81_groups_0, pad = query_81_pad_0, pad_type = query_81_pad_type_0, strides = query_81_strides_0, weight = layers_20_self_attn_q_proj_weight_to_fp16, x = obj_83_cast_fp16)[name = tensor<string, []>("query_81_cast_fp16")];
             tensor<string, []> key_41_pad_type_0 = const()[name = tensor<string, []>("key_41_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> key_41_strides_0 = const()[name = tensor<string, []>("key_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> key_41_pad_0 = const()[name = tensor<string, []>("key_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
@@ -3712,173 +3714,173 @@ program(1.0)
             tensor<int32, [2]> value_41_dilations_0 = const()[name = tensor<string, []>("value_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> value_41_groups_0 = const()[name = tensor<string, []>("value_41_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_20_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_20_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1037996864)))];
-            tensor<fp16, [1, 1024, 1, 188]> value_41_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = value_41_dilations_0, groups = value_41_groups_0, pad = value_41_pad_0, pad_type = value_41_pad_type_0, strides = value_41_strides_0, weight = layers_20_self_attn_v_proj_weight_to_fp16, x = obj_83_cast_fp16)[name = tensor<string, []>("value_41_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_5880_to_fp16 = const()[name = tensor<string, []>("op_5880_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1040094080)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_83_cast_fp16 = add(x = query_81_cast_fp16, y = var_5880_to_fp16)[name = tensor<string, []>("query_83_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_5883_to_fp16 = const()[name = tensor<string, []>("op_5883_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1040096192)))];
-            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_41_cast_fp16 = add(x = query_81_cast_fp16, y = var_5883_to_fp16)[name = tensor<string, []>("q_with_bias_v_41_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> value_41_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_41_dilations_0, groups = value_41_groups_0, pad = value_41_pad_0, pad_type = value_41_pad_type_0, strides = value_41_strides_0, weight = layers_20_self_attn_v_proj_weight_to_fp16, x = obj_83_cast_fp16)[name = tensor<string, []>("value_41_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_5878_to_fp16 = const()[name = tensor<string, []>("op_5878_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1040094080)))];
+            tensor<fp16, [1, 1024, 1, 188]> query_83_cast_fp16 = add(x = query_81_cast_fp16, y = var_5878_to_fp16)[name = tensor<string, []>("query_83_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_5881_to_fp16 = const()[name = tensor<string, []>("op_5881_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1040096192)))];
+            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_41_cast_fp16 = add(x = query_81_cast_fp16, y = var_5881_to_fp16)[name = tensor<string, []>("q_with_bias_v_41_cast_fp16")];
             tensor<string, []> p_41_pad_type_0 = const()[name = tensor<string, []>("p_41_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> p_41_strides_0 = const()[name = tensor<string, []>("p_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> p_41_pad_0 = const()[name = tensor<string, []>("p_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> p_41_dilations_0 = const()[name = tensor<string, []>("p_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> p_41_groups_0 = const()[name = tensor<string, []>("p_41_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_20_self_attn_linear_pos_weight_to_fp16 = const()[name = tensor<string, []>("layers_20_self_attn_linear_pos_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1040098304)))];
-            tensor<fp16, [1, 1024, 1, 375]> p_41_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = p_41_dilations_0, groups = p_41_groups_0, pad = p_41_pad_0, pad_type = p_41_pad_type_0, strides = p_41_strides_0, weight = layers_20_self_attn_linear_pos_weight_to_fp16, x = pos_enc_to_fp16)[name = tensor<string, []>("p_41_cast_fp16")];
-            tensor<int32, [4]> var_5894 = const()[name = tensor<string, []>("op_5894"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_5895_cast_fp16 = reshape(shape = var_5894, x = q_with_bias_v_41_cast_fp16)[name = tensor<string, []>("op_5895_cast_fp16")];
-            tensor<int32, [4]> var_5896 = const()[name = tensor<string, []>("op_5896"), val = tensor<int32, [4]>([1, 8, 128, -1])];
-            tensor<fp16, [1, 8, 128, 375]> var_5897_cast_fp16 = reshape(shape = var_5896, x = p_41_cast_fp16)[name = tensor<string, []>("op_5897_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 375]> p_41_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_41_dilations_0, groups = p_41_groups_0, pad = p_41_pad_0, pad_type = p_41_pad_type_0, strides = p_41_strides_0, weight = layers_20_self_attn_linear_pos_weight_to_fp16, x = obj_3_cast_fp16)[name = tensor<string, []>("p_41_cast_fp16")];
+            tensor<int32, [4]> var_5892 = const()[name = tensor<string, []>("op_5892"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_5893_cast_fp16 = reshape(shape = var_5892, x = q_with_bias_v_41_cast_fp16)[name = tensor<string, []>("op_5893_cast_fp16")];
+            tensor<int32, [4]> var_5894 = const()[name = tensor<string, []>("op_5894"), val = tensor<int32, [4]>([1, 8, 128, -1])];
+            tensor<fp16, [1, 8, 128, 375]> var_5895_cast_fp16 = reshape(shape = var_5894, x = p_41_cast_fp16)[name = tensor<string, []>("op_5895_cast_fp16")];
             tensor<bool, []> matrix_bd_161_transpose_x_0 = const()[name = tensor<string, []>("matrix_bd_161_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> matrix_bd_161_transpose_y_0 = const()[name = tensor<string, []>("matrix_bd_161_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_161_cast_fp16 = matmul(transpose_x = matrix_bd_161_transpose_x_0, transpose_y = matrix_bd_161_transpose_y_0, x = var_5895_cast_fp16, y = var_5897_cast_fp16)[name = tensor<string, []>("matrix_bd_161_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_161_cast_fp16 = matmul(transpose_x = matrix_bd_161_transpose_x_0, transpose_y = matrix_bd_161_transpose_y_0, x = var_5893_cast_fp16, y = var_5895_cast_fp16)[name = tensor<string, []>("matrix_bd_161_cast_fp16")];
             tensor<int32, [8]> matrix_bd_163_pad_0 = const()[name = tensor<string, []>("matrix_bd_163_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
             tensor<string, []> matrix_bd_163_mode_0 = const()[name = tensor<string, []>("matrix_bd_163_mode_0"), val = tensor<string, []>("constant")];
             tensor<fp16, []> const_230_to_fp16 = const()[name = tensor<string, []>("const_230_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
             tensor<fp16, [1, 8, 188, 376]> matrix_bd_163_cast_fp16 = pad(constant_val = const_230_to_fp16, mode = matrix_bd_163_mode_0, pad = matrix_bd_163_pad_0, x = matrix_bd_161_cast_fp16)[name = tensor<string, []>("matrix_bd_163_cast_fp16")];
-            tensor<int32, [4]> var_5906 = const()[name = tensor<string, []>("op_5906"), val = tensor<int32, [4]>([1, 8, -1, 188])];
-            tensor<fp16, [1, 8, 376, 188]> matrix_bd_165_cast_fp16 = reshape(shape = var_5906, x = matrix_bd_163_cast_fp16)[name = tensor<string, []>("matrix_bd_165_cast_fp16")];
-            tensor<int32, [4]> var_5910_begin_0 = const()[name = tensor<string, []>("op_5910_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
-            tensor<int32, [4]> var_5910_end_0 = const()[name = tensor<string, []>("op_5910_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
-            tensor<bool, [4]> var_5910_end_mask_0 = const()[name = tensor<string, []>("op_5910_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
-            tensor<fp16, [1, 8, 375, 188]> var_5910_cast_fp16 = slice_by_index(begin = var_5910_begin_0, end = var_5910_end_0, end_mask = var_5910_end_mask_0, x = matrix_bd_165_cast_fp16)[name = tensor<string, []>("op_5910_cast_fp16")];
-            tensor<int32, [4]> var_5911 = const()[name = tensor<string, []>("op_5911"), val = tensor<int32, [4]>([1, 8, 188, 375])];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_167_cast_fp16 = reshape(shape = var_5911, x = var_5910_cast_fp16)[name = tensor<string, []>("matrix_bd_167_cast_fp16")];
-            tensor<int32, [4]> var_5916_begin_0 = const()[name = tensor<string, []>("op_5916_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [4]> var_5916_end_0 = const()[name = tensor<string, []>("op_5916_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
-            tensor<bool, [4]> var_5916_end_mask_0 = const()[name = tensor<string, []>("op_5916_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
-            tensor<fp16, [1, 8, 188, 188]> var_5916_cast_fp16 = slice_by_index(begin = var_5916_begin_0, end = var_5916_end_0, end_mask = var_5916_end_mask_0, x = matrix_bd_167_cast_fp16)[name = tensor<string, []>("op_5916_cast_fp16")];
-            tensor<fp16, []> var_5917_to_fp16 = const()[name = tensor<string, []>("op_5917_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 188, 188]> qk_mask_41_cast_fp16 = mul(x = var_5916_cast_fp16, y = var_5917_to_fp16)[name = tensor<string, []>("qk_mask_41_cast_fp16")];
-            tensor<int32, [4]> var_5921 = const()[name = tensor<string, []>("op_5921"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> mh_q_41_cast_fp16 = reshape(shape = var_5921, x = query_83_cast_fp16)[name = tensor<string, []>("mh_q_41_cast_fp16")];
-            tensor<fp16, []> var_5923_to_fp16 = const()[name = tensor<string, []>("op_5923_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 128, 188]> var_5924_cast_fp16 = mul(x = mh_q_41_cast_fp16, y = var_5923_to_fp16)[name = tensor<string, []>("op_5924_cast_fp16")];
-            tensor<int32, [4]> var_5927 = const()[name = tensor<string, []>("op_5927"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_5928_cast_fp16 = reshape(shape = var_5927, x = key_41_cast_fp16)[name = tensor<string, []>("op_5928_cast_fp16")];
+            tensor<int32, [4]> var_5904 = const()[name = tensor<string, []>("op_5904"), val = tensor<int32, [4]>([1, 8, -1, 188])];
+            tensor<fp16, [1, 8, 376, 188]> matrix_bd_165_cast_fp16 = reshape(shape = var_5904, x = matrix_bd_163_cast_fp16)[name = tensor<string, []>("matrix_bd_165_cast_fp16")];
+            tensor<int32, [4]> var_5908_begin_0 = const()[name = tensor<string, []>("op_5908_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_5908_end_0 = const()[name = tensor<string, []>("op_5908_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
+            tensor<bool, [4]> var_5908_end_mask_0 = const()[name = tensor<string, []>("op_5908_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 375, 188]> var_5908_cast_fp16 = slice_by_index(begin = var_5908_begin_0, end = var_5908_end_0, end_mask = var_5908_end_mask_0, x = matrix_bd_165_cast_fp16)[name = tensor<string, []>("op_5908_cast_fp16")];
+            tensor<int32, [4]> var_5909 = const()[name = tensor<string, []>("op_5909"), val = tensor<int32, [4]>([1, 8, 188, 375])];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_167_cast_fp16 = reshape(shape = var_5909, x = var_5908_cast_fp16)[name = tensor<string, []>("matrix_bd_167_cast_fp16")];
+            tensor<int32, [4]> var_5914_begin_0 = const()[name = tensor<string, []>("op_5914_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_5914_end_0 = const()[name = tensor<string, []>("op_5914_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
+            tensor<bool, [4]> var_5914_end_mask_0 = const()[name = tensor<string, []>("op_5914_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 188, 188]> var_5914_cast_fp16 = slice_by_index(begin = var_5914_begin_0, end = var_5914_end_0, end_mask = var_5914_end_mask_0, x = matrix_bd_167_cast_fp16)[name = tensor<string, []>("op_5914_cast_fp16")];
+            tensor<fp16, []> var_5915_to_fp16 = const()[name = tensor<string, []>("op_5915_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 188, 188]> qk_mask_41_cast_fp16 = mul(x = var_5914_cast_fp16, y = var_5915_to_fp16)[name = tensor<string, []>("qk_mask_41_cast_fp16")];
+            tensor<int32, [4]> var_5919 = const()[name = tensor<string, []>("op_5919"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> mh_q_41_cast_fp16 = reshape(shape = var_5919, x = query_83_cast_fp16)[name = tensor<string, []>("mh_q_41_cast_fp16")];
+            tensor<fp16, []> var_5921_to_fp16 = const()[name = tensor<string, []>("op_5921_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 128, 188]> var_5922_cast_fp16 = mul(x = mh_q_41_cast_fp16, y = var_5921_to_fp16)[name = tensor<string, []>("op_5922_cast_fp16")];
+            tensor<int32, [4]> var_5925 = const()[name = tensor<string, []>("op_5925"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_5926_cast_fp16 = reshape(shape = var_5925, x = key_41_cast_fp16)[name = tensor<string, []>("op_5926_cast_fp16")];
             tensor<bool, []> mh_w_81_transpose_x_0 = const()[name = tensor<string, []>("mh_w_81_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> mh_w_81_transpose_y_0 = const()[name = tensor<string, []>("mh_w_81_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 188]> mh_w_81_cast_fp16 = matmul(transpose_x = mh_w_81_transpose_x_0, transpose_y = mh_w_81_transpose_y_0, x = var_5924_cast_fp16, y = var_5928_cast_fp16)[name = tensor<string, []>("mh_w_81_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> mh_w_81_cast_fp16 = matmul(transpose_x = mh_w_81_transpose_x_0, transpose_y = mh_w_81_transpose_y_0, x = var_5922_cast_fp16, y = var_5926_cast_fp16)[name = tensor<string, []>("mh_w_81_cast_fp16")];
             tensor<fp16, [1, 8, 188, 188]> mh_w_83_cast_fp16 = add(x = mh_w_81_cast_fp16, y = qk_mask_41_cast_fp16)[name = tensor<string, []>("mh_w_83_cast_fp16")];
-            tensor<fp16, [1, 8, 188, 188]> var_5932_cast_fp16 = softmax(axis = var_5773, x = mh_w_83_cast_fp16)[name = tensor<string, []>("op_5932_cast_fp16")];
-            tensor<int32, [4]> var_5933 = const()[name = tensor<string, []>("op_5933"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_5934_cast_fp16 = reshape(shape = var_5933, x = value_41_cast_fp16)[name = tensor<string, []>("op_5934_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> var_5930_cast_fp16 = softmax(axis = var_5771, x = mh_w_83_cast_fp16)[name = tensor<string, []>("op_5930_cast_fp16")];
+            tensor<int32, [4]> var_5931 = const()[name = tensor<string, []>("op_5931"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_5932_cast_fp16 = reshape(shape = var_5931, x = value_41_cast_fp16)[name = tensor<string, []>("op_5932_cast_fp16")];
             tensor<bool, []> attn_41_transpose_x_0 = const()[name = tensor<string, []>("attn_41_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> attn_41_transpose_y_0 = const()[name = tensor<string, []>("attn_41_transpose_y_0"), val = tensor<bool, []>(true)];
-            tensor<fp16, [1, 8, 128, 188]> attn_41_cast_fp16 = matmul(transpose_x = attn_41_transpose_x_0, transpose_y = attn_41_transpose_y_0, x = var_5934_cast_fp16, y = var_5932_cast_fp16)[name = tensor<string, []>("attn_41_cast_fp16")];
-            tensor<int32, [4]> var_5937 = const()[name = tensor<string, []>("op_5937"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
-            tensor<fp16, [1, 1024, 1, 188]> input_541_cast_fp16 = reshape(shape = var_5937, x = attn_41_cast_fp16)[name = tensor<string, []>("input_541_cast_fp16")];
+            tensor<fp16, [1, 8, 128, 188]> attn_41_cast_fp16 = matmul(transpose_x = attn_41_transpose_x_0, transpose_y = attn_41_transpose_y_0, x = var_5932_cast_fp16, y = var_5930_cast_fp16)[name = tensor<string, []>("attn_41_cast_fp16")];
+            tensor<int32, [4]> var_5935 = const()[name = tensor<string, []>("op_5935"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
+            tensor<fp16, [1, 1024, 1, 188]> input_543_cast_fp16 = reshape(shape = var_5935, x = attn_41_cast_fp16)[name = tensor<string, []>("input_543_cast_fp16")];
             tensor<string, []> obj_85_pad_type_0 = const()[name = tensor<string, []>("obj_85_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> obj_85_strides_0 = const()[name = tensor<string, []>("obj_85_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> obj_85_pad_0 = const()[name = tensor<string, []>("obj_85_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> obj_85_dilations_0 = const()[name = tensor<string, []>("obj_85_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> obj_85_groups_0 = const()[name = tensor<string, []>("obj_85_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_20_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_20_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1042195520)))];
-            tensor<fp16, [1, 1024, 1, 188]> obj_85_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = obj_85_dilations_0, groups = obj_85_groups_0, pad = obj_85_pad_0, pad_type = obj_85_pad_type_0, strides = obj_85_strides_0, weight = layers_20_self_attn_o_proj_weight_to_fp16, x = input_541_cast_fp16)[name = tensor<string, []>("obj_85_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_85_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_85_dilations_0, groups = obj_85_groups_0, pad = obj_85_pad_0, pad_type = obj_85_pad_type_0, strides = obj_85_strides_0, weight = layers_20_self_attn_o_proj_weight_to_fp16, x = input_543_cast_fp16)[name = tensor<string, []>("obj_85_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_205_cast_fp16 = add(x = inputs_203_cast_fp16, y = obj_85_cast_fp16)[name = tensor<string, []>("inputs_205_cast_fp16")];
             tensor<int32, [1]> out_205_axes_0 = const()[name = tensor<string, []>("out_205_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_5955_to_fp16 = const()[name = tensor<string, []>("op_5955_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_205_cast_fp16 = layer_norm(axes = out_205_axes_0, epsilon = var_5955_to_fp16, x = inputs_205_cast_fp16)[name = tensor<string, []>("out_205_cast_fp16")];
-            tensor<fp16, [1024]> input_543_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_543_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1044292736)))];
-            tensor<fp16, [1024]> input_543_beta_0_to_fp16 = const()[name = tensor<string, []>("input_543_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1044294848)))];
-            tensor<fp16, []> input_543_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_543_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_543_cast_fp16 = batch_norm(beta = input_543_beta_0_to_fp16, epsilon = input_543_epsilon_0_to_fp16, gamma = input_543_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_205_cast_fp16)[name = tensor<string, []>("input_543_cast_fp16")];
-            tensor<string, []> input_545_pad_type_0 = const()[name = tensor<string, []>("input_545_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_545_strides_0 = const()[name = tensor<string, []>("input_545_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_545_pad_0 = const()[name = tensor<string, []>("input_545_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_545_dilations_0 = const()[name = tensor<string, []>("input_545_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_545_groups_0 = const()[name = tensor<string, []>("input_545_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_5953_to_fp16 = const()[name = tensor<string, []>("op_5953_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_205_cast_fp16 = layer_norm(axes = out_205_axes_0, epsilon = var_5953_to_fp16, x = inputs_205_cast_fp16)[name = tensor<string, []>("out_205_cast_fp16")];
+            tensor<fp16, [1024]> input_545_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_545_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1044292736)))];
+            tensor<fp16, [1024]> input_545_beta_0_to_fp16 = const()[name = tensor<string, []>("input_545_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1044294848)))];
+            tensor<fp16, []> input_545_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_545_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_545_cast_fp16 = batch_norm(beta = input_545_beta_0_to_fp16, epsilon = input_545_epsilon_0_to_fp16, gamma = input_545_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_205_cast_fp16)[name = tensor<string, []>("input_545_cast_fp16")];
+            tensor<string, []> input_547_pad_type_0 = const()[name = tensor<string, []>("input_547_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_547_strides_0 = const()[name = tensor<string, []>("input_547_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_547_pad_0 = const()[name = tensor<string, []>("input_547_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_547_dilations_0 = const()[name = tensor<string, []>("input_547_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_547_groups_0 = const()[name = tensor<string, []>("input_547_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [2048, 1024, 1, 1]> layers_20_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("layers_20_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1044296960)))];
-            tensor<fp16, [1, 2048, 1, 188]> input_545_cast_fp16 = conv(dilations = input_545_dilations_0, groups = input_545_groups_0, pad = input_545_pad_0, pad_type = input_545_pad_type_0, strides = input_545_strides_0, weight = layers_20_conv_pointwise_conv1_weight_to_fp16, x = input_543_cast_fp16)[name = tensor<string, []>("input_545_cast_fp16")];
-            tensor<int32, []> input_547_split_num_splits_0 = const()[name = tensor<string, []>("input_547_split_num_splits_0"), val = tensor<int32, []>(2)];
-            tensor<int32, []> input_547_split_axis_0 = const()[name = tensor<string, []>("input_547_split_axis_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1, 1024, 1, 188]> input_547_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_547_split_cast_fp16_1 = split(axis = input_547_split_axis_0, num_splits = input_547_split_num_splits_0, x = input_545_cast_fp16)[name = tensor<string, []>("input_547_split_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_547_split_1_sigmoid_cast_fp16 = sigmoid(x = input_547_split_cast_fp16_1)[name = tensor<string, []>("input_547_split_1_sigmoid_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_547_cast_fp16 = mul(x = input_547_split_cast_fp16_0, y = input_547_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_547_cast_fp16")];
-            tensor<string, []> input_549_pad_type_0 = const()[name = tensor<string, []>("input_549_pad_type_0"), val = tensor<string, []>("custom")];
-            tensor<int32, [4]> input_549_pad_0 = const()[name = tensor<string, []>("input_549_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
-            tensor<int32, []> input_549_groups_0 = const()[name = tensor<string, []>("input_549_groups_0"), val = tensor<int32, []>(1024)];
-            tensor<int32, [2]> input_549_strides_0 = const()[name = tensor<string, []>("input_549_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> input_549_dilations_0 = const()[name = tensor<string, []>("input_549_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [1, 2048, 1, 188]> input_547_cast_fp16 = conv(dilations = input_547_dilations_0, groups = input_547_groups_0, pad = input_547_pad_0, pad_type = input_547_pad_type_0, strides = input_547_strides_0, weight = layers_20_conv_pointwise_conv1_weight_to_fp16, x = input_545_cast_fp16)[name = tensor<string, []>("input_547_cast_fp16")];
+            tensor<int32, []> input_549_split_num_splits_0 = const()[name = tensor<string, []>("input_549_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> input_549_split_axis_0 = const()[name = tensor<string, []>("input_549_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 1, 188]> input_549_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_549_split_cast_fp16_1 = split(axis = input_549_split_axis_0, num_splits = input_549_split_num_splits_0, x = input_547_cast_fp16)[name = tensor<string, []>("input_549_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_549_split_1_sigmoid_cast_fp16 = sigmoid(x = input_549_split_cast_fp16_1)[name = tensor<string, []>("input_549_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_549_cast_fp16 = mul(x = input_549_split_cast_fp16_0, y = input_549_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_549_cast_fp16")];
+            tensor<string, []> input_551_pad_type_0 = const()[name = tensor<string, []>("input_551_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_551_pad_0 = const()[name = tensor<string, []>("input_551_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
+            tensor<int32, []> input_551_groups_0 = const()[name = tensor<string, []>("input_551_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [2]> input_551_strides_0 = const()[name = tensor<string, []>("input_551_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> input_551_dilations_0 = const()[name = tensor<string, []>("input_551_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<fp16, [1024, 1, 1, 9]> const_308_to_fp16 = const()[name = tensor<string, []>("const_308_to_fp16"), val = tensor<fp16, [1024, 1, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1048491328)))];
             tensor<fp16, [1024]> const_309_to_fp16 = const()[name = tensor<string, []>("const_309_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1048509824)))];
-            tensor<fp16, [1, 1024, 1, 188]> input_551_cast_fp16 = conv(bias = const_309_to_fp16, dilations = input_549_dilations_0, groups = input_549_groups_0, pad = input_549_pad_0, pad_type = input_549_pad_type_0, strides = input_549_strides_0, weight = const_308_to_fp16, x = input_547_cast_fp16)[name = tensor<string, []>("input_551_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_553_cast_fp16 = silu(x = input_551_cast_fp16)[name = tensor<string, []>("input_553_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_553_cast_fp16 = conv(bias = const_309_to_fp16, dilations = input_551_dilations_0, groups = input_551_groups_0, pad = input_551_pad_0, pad_type = input_551_pad_type_0, strides = input_551_strides_0, weight = const_308_to_fp16, x = input_549_cast_fp16)[name = tensor<string, []>("input_553_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_555_cast_fp16 = silu(x = input_553_cast_fp16)[name = tensor<string, []>("input_555_cast_fp16")];
             tensor<string, []> x_125_pad_type_0 = const()[name = tensor<string, []>("x_125_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_125_strides_0 = const()[name = tensor<string, []>("x_125_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_125_pad_0 = const()[name = tensor<string, []>("x_125_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_125_dilations_0 = const()[name = tensor<string, []>("x_125_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_125_groups_0 = const()[name = tensor<string, []>("x_125_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_20_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("layers_20_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1048511936)))];
-            tensor<fp16, [1, 1024, 1, 188]> x_125_cast_fp16 = conv(dilations = x_125_dilations_0, groups = x_125_groups_0, pad = x_125_pad_0, pad_type = x_125_pad_type_0, strides = x_125_strides_0, weight = layers_20_conv_pointwise_conv2_weight_to_fp16, x = input_553_cast_fp16)[name = tensor<string, []>("x_125_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> x_125_cast_fp16 = conv(dilations = x_125_dilations_0, groups = x_125_groups_0, pad = x_125_pad_0, pad_type = x_125_pad_type_0, strides = x_125_strides_0, weight = layers_20_conv_pointwise_conv2_weight_to_fp16, x = input_555_cast_fp16)[name = tensor<string, []>("x_125_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_207_cast_fp16 = add(x = inputs_205_cast_fp16, y = x_125_cast_fp16)[name = tensor<string, []>("inputs_207_cast_fp16")];
             tensor<int32, [1]> out_207_axes_0 = const()[name = tensor<string, []>("out_207_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_6003_to_fp16 = const()[name = tensor<string, []>("op_6003_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_207_cast_fp16 = layer_norm(axes = out_207_axes_0, epsilon = var_6003_to_fp16, x = inputs_207_cast_fp16)[name = tensor<string, []>("out_207_cast_fp16")];
-            tensor<fp16, [1024]> input_555_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_555_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1050609152)))];
-            tensor<fp16, [1024]> input_555_beta_0_to_fp16 = const()[name = tensor<string, []>("input_555_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1050611264)))];
-            tensor<fp16, []> input_555_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_555_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_555_cast_fp16 = batch_norm(beta = input_555_beta_0_to_fp16, epsilon = input_555_epsilon_0_to_fp16, gamma = input_555_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_207_cast_fp16)[name = tensor<string, []>("input_555_cast_fp16")];
-            tensor<string, []> input_557_pad_type_0 = const()[name = tensor<string, []>("input_557_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_557_strides_0 = const()[name = tensor<string, []>("input_557_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_557_pad_0 = const()[name = tensor<string, []>("input_557_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_557_dilations_0 = const()[name = tensor<string, []>("input_557_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_557_groups_0 = const()[name = tensor<string, []>("input_557_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_6001_to_fp16 = const()[name = tensor<string, []>("op_6001_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_207_cast_fp16 = layer_norm(axes = out_207_axes_0, epsilon = var_6001_to_fp16, x = inputs_207_cast_fp16)[name = tensor<string, []>("out_207_cast_fp16")];
+            tensor<fp16, [1024]> input_557_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_557_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1050609152)))];
+            tensor<fp16, [1024]> input_557_beta_0_to_fp16 = const()[name = tensor<string, []>("input_557_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1050611264)))];
+            tensor<fp16, []> input_557_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_557_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_557_cast_fp16 = batch_norm(beta = input_557_beta_0_to_fp16, epsilon = input_557_epsilon_0_to_fp16, gamma = input_557_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_207_cast_fp16)[name = tensor<string, []>("input_557_cast_fp16")];
+            tensor<string, []> input_559_pad_type_0 = const()[name = tensor<string, []>("input_559_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_559_strides_0 = const()[name = tensor<string, []>("input_559_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_559_pad_0 = const()[name = tensor<string, []>("input_559_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_559_dilations_0 = const()[name = tensor<string, []>("input_559_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_559_groups_0 = const()[name = tensor<string, []>("input_559_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_20_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_20_feed_forward2_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1050613376)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_557_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_557_dilations_0, groups = input_557_groups_0, pad = input_557_pad_0, pad_type = input_557_pad_type_0, strides = input_557_strides_0, weight = layers_20_feed_forward2_fc1_weight_to_fp16, x = input_555_cast_fp16)[name = tensor<string, []>("input_557_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_559_cast_fp16 = silu(x = input_557_cast_fp16)[name = tensor<string, []>("input_559_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_559_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_559_dilations_0, groups = input_559_groups_0, pad = input_559_pad_0, pad_type = input_559_pad_type_0, strides = input_559_strides_0, weight = layers_20_feed_forward2_fc1_weight_to_fp16, x = input_557_cast_fp16)[name = tensor<string, []>("input_559_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_561_cast_fp16 = silu(x = input_559_cast_fp16)[name = tensor<string, []>("input_561_cast_fp16")];
             tensor<string, []> x_127_pad_type_0 = const()[name = tensor<string, []>("x_127_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_127_strides_0 = const()[name = tensor<string, []>("x_127_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_127_pad_0 = const()[name = tensor<string, []>("x_127_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_127_dilations_0 = const()[name = tensor<string, []>("x_127_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_127_groups_0 = const()[name = tensor<string, []>("x_127_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_6031_weight_0_to_fp16 = const()[name = tensor<string, []>("op_6031_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1059002048)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_6031_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_127_dilations_0, groups = x_127_groups_0, pad = x_127_pad_0, pad_type = x_127_pad_type_0, strides = x_127_strides_0, weight = var_6031_weight_0_to_fp16, x = input_559_cast_fp16)[name = tensor<string, []>("op_6031_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_209_cast_fp16 = add(x = inputs_207_cast_fp16, y = var_6031_cast_fp16)[name = tensor<string, []>("inputs_209_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_6029_weight_0_to_fp16 = const()[name = tensor<string, []>("op_6029_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1059002048)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_6029_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_127_dilations_0, groups = x_127_groups_0, pad = x_127_pad_0, pad_type = x_127_pad_type_0, strides = x_127_strides_0, weight = var_6029_weight_0_to_fp16, x = input_561_cast_fp16)[name = tensor<string, []>("op_6029_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_209_cast_fp16 = add(x = inputs_207_cast_fp16, y = var_6029_cast_fp16)[name = tensor<string, []>("inputs_209_cast_fp16")];
             tensor<int32, [1]> out_209_axes_0 = const()[name = tensor<string, []>("out_209_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_6041_to_fp16 = const()[name = tensor<string, []>("op_6041_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_209_cast_fp16 = layer_norm(axes = out_209_axes_0, epsilon = var_6041_to_fp16, x = inputs_209_cast_fp16)[name = tensor<string, []>("out_209_cast_fp16")];
+            tensor<fp16, []> var_6039_to_fp16 = const()[name = tensor<string, []>("op_6039_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_209_cast_fp16 = layer_norm(axes = out_209_axes_0, epsilon = var_6039_to_fp16, x = inputs_209_cast_fp16)[name = tensor<string, []>("out_209_cast_fp16")];
             tensor<fp16, [1024]> inputs_211_gamma_0_to_fp16 = const()[name = tensor<string, []>("inputs_211_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1067390720)))];
             tensor<fp16, [1024]> inputs_211_beta_0_to_fp16 = const()[name = tensor<string, []>("inputs_211_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1067392832)))];
             tensor<fp16, []> inputs_211_epsilon_0_to_fp16 = const()[name = tensor<string, []>("inputs_211_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_211_cast_fp16 = batch_norm(beta = inputs_211_beta_0_to_fp16, epsilon = inputs_211_epsilon_0_to_fp16, gamma = inputs_211_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_209_cast_fp16)[name = tensor<string, []>("inputs_211_cast_fp16")];
-            tensor<int32, []> var_6055 = const()[name = tensor<string, []>("op_6055"), val = tensor<int32, []>(3)];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_211_cast_fp16 = batch_norm(beta = inputs_211_beta_0_to_fp16, epsilon = inputs_211_epsilon_0_to_fp16, gamma = inputs_211_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_209_cast_fp16)[name = tensor<string, []>("inputs_211_cast_fp16")];
+            tensor<int32, []> var_6053 = const()[name = tensor<string, []>("op_6053"), val = tensor<int32, []>(3)];
             tensor<int32, [1]> out_211_axes_0 = const()[name = tensor<string, []>("out_211_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_6086_to_fp16 = const()[name = tensor<string, []>("op_6086_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_211_cast_fp16 = layer_norm(axes = out_211_axes_0, epsilon = var_6086_to_fp16, x = inputs_211_cast_fp16)[name = tensor<string, []>("out_211_cast_fp16")];
-            tensor<fp16, [1024]> input_561_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_561_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1067394944)))];
-            tensor<fp16, [1024]> input_561_beta_0_to_fp16 = const()[name = tensor<string, []>("input_561_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1067397056)))];
-            tensor<fp16, []> input_561_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_561_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_561_cast_fp16 = batch_norm(beta = input_561_beta_0_to_fp16, epsilon = input_561_epsilon_0_to_fp16, gamma = input_561_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_211_cast_fp16)[name = tensor<string, []>("input_561_cast_fp16")];
-            tensor<string, []> input_563_pad_type_0 = const()[name = tensor<string, []>("input_563_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_563_strides_0 = const()[name = tensor<string, []>("input_563_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_563_pad_0 = const()[name = tensor<string, []>("input_563_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_563_dilations_0 = const()[name = tensor<string, []>("input_563_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_563_groups_0 = const()[name = tensor<string, []>("input_563_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_6084_to_fp16 = const()[name = tensor<string, []>("op_6084_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_211_cast_fp16 = layer_norm(axes = out_211_axes_0, epsilon = var_6084_to_fp16, x = inputs_211_cast_fp16)[name = tensor<string, []>("out_211_cast_fp16")];
+            tensor<fp16, [1024]> input_563_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_563_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1067394944)))];
+            tensor<fp16, [1024]> input_563_beta_0_to_fp16 = const()[name = tensor<string, []>("input_563_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1067397056)))];
+            tensor<fp16, []> input_563_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_563_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_563_cast_fp16 = batch_norm(beta = input_563_beta_0_to_fp16, epsilon = input_563_epsilon_0_to_fp16, gamma = input_563_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_211_cast_fp16)[name = tensor<string, []>("input_563_cast_fp16")];
+            tensor<string, []> input_565_pad_type_0 = const()[name = tensor<string, []>("input_565_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_565_strides_0 = const()[name = tensor<string, []>("input_565_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_565_pad_0 = const()[name = tensor<string, []>("input_565_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_565_dilations_0 = const()[name = tensor<string, []>("input_565_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_565_groups_0 = const()[name = tensor<string, []>("input_565_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_21_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_21_feed_forward1_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1067399168)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_563_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_563_dilations_0, groups = input_563_groups_0, pad = input_563_pad_0, pad_type = input_563_pad_type_0, strides = input_563_strides_0, weight = layers_21_feed_forward1_fc1_weight_to_fp16, x = input_561_cast_fp16)[name = tensor<string, []>("input_563_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_565_cast_fp16 = silu(x = input_563_cast_fp16)[name = tensor<string, []>("input_565_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_565_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_565_dilations_0, groups = input_565_groups_0, pad = input_565_pad_0, pad_type = input_565_pad_type_0, strides = input_565_strides_0, weight = layers_21_feed_forward1_fc1_weight_to_fp16, x = input_563_cast_fp16)[name = tensor<string, []>("input_565_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_567_cast_fp16 = silu(x = input_565_cast_fp16)[name = tensor<string, []>("input_567_cast_fp16")];
             tensor<string, []> x_129_pad_type_0 = const()[name = tensor<string, []>("x_129_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_129_strides_0 = const()[name = tensor<string, []>("x_129_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_129_pad_0 = const()[name = tensor<string, []>("x_129_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_129_dilations_0 = const()[name = tensor<string, []>("x_129_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_129_groups_0 = const()[name = tensor<string, []>("x_129_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_6114_weight_0_to_fp16 = const()[name = tensor<string, []>("op_6114_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1075787840)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_6114_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_129_dilations_0, groups = x_129_groups_0, pad = x_129_pad_0, pad_type = x_129_pad_type_0, strides = x_129_strides_0, weight = var_6114_weight_0_to_fp16, x = input_565_cast_fp16)[name = tensor<string, []>("op_6114_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_213_cast_fp16 = add(x = inputs_211_cast_fp16, y = var_6114_cast_fp16)[name = tensor<string, []>("inputs_213_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_6112_weight_0_to_fp16 = const()[name = tensor<string, []>("op_6112_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1075787840)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_6112_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_129_dilations_0, groups = x_129_groups_0, pad = x_129_pad_0, pad_type = x_129_pad_type_0, strides = x_129_strides_0, weight = var_6112_weight_0_to_fp16, x = input_567_cast_fp16)[name = tensor<string, []>("op_6112_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_213_cast_fp16 = add(x = inputs_211_cast_fp16, y = var_6112_cast_fp16)[name = tensor<string, []>("inputs_213_cast_fp16")];
             tensor<int32, [1]> out_213_axes_0 = const()[name = tensor<string, []>("out_213_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_6124_to_fp16 = const()[name = tensor<string, []>("op_6124_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_213_cast_fp16 = layer_norm(axes = out_213_axes_0, epsilon = var_6124_to_fp16, x = inputs_213_cast_fp16)[name = tensor<string, []>("out_213_cast_fp16")];
+            tensor<fp16, []> var_6122_to_fp16 = const()[name = tensor<string, []>("op_6122_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_213_cast_fp16 = layer_norm(axes = out_213_axes_0, epsilon = var_6122_to_fp16, x = inputs_213_cast_fp16)[name = tensor<string, []>("out_213_cast_fp16")];
             tensor<fp16, [1024]> obj_87_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_87_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1084176512)))];
             tensor<fp16, [1024]> obj_87_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_87_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1084178624)))];
             tensor<fp16, []> obj_87_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_87_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> obj_87_cast_fp16 = batch_norm(beta = obj_87_beta_0_to_fp16, epsilon = obj_87_epsilon_0_to_fp16, gamma = obj_87_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_213_cast_fp16)[name = tensor<string, []>("obj_87_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_87_cast_fp16 = batch_norm(beta = obj_87_beta_0_to_fp16, epsilon = obj_87_epsilon_0_to_fp16, gamma = obj_87_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_213_cast_fp16)[name = tensor<string, []>("obj_87_cast_fp16")];
             tensor<string, []> query_85_pad_type_0 = const()[name = tensor<string, []>("query_85_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> query_85_strides_0 = const()[name = tensor<string, []>("query_85_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> query_85_pad_0 = const()[name = tensor<string, []>("query_85_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> query_85_dilations_0 = const()[name = tensor<string, []>("query_85_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> query_85_groups_0 = const()[name = tensor<string, []>("query_85_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_21_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_21_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1084180736)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_85_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = query_85_dilations_0, groups = query_85_groups_0, pad = query_85_pad_0, pad_type = query_85_pad_type_0, strides = query_85_strides_0, weight = layers_21_self_attn_q_proj_weight_to_fp16, x = obj_87_cast_fp16)[name = tensor<string, []>("query_85_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> query_85_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_85_dilations_0, groups = query_85_groups_0, pad = query_85_pad_0, pad_type = query_85_pad_type_0, strides = query_85_strides_0, weight = layers_21_self_attn_q_proj_weight_to_fp16, x = obj_87_cast_fp16)[name = tensor<string, []>("query_85_cast_fp16")];
             tensor<string, []> key_43_pad_type_0 = const()[name = tensor<string, []>("key_43_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> key_43_strides_0 = const()[name = tensor<string, []>("key_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> key_43_pad_0 = const()[name = tensor<string, []>("key_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
@@ -3892,173 +3894,173 @@ program(1.0)
             tensor<int32, [2]> value_43_dilations_0 = const()[name = tensor<string, []>("value_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> value_43_groups_0 = const()[name = tensor<string, []>("value_43_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_21_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_21_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1088375168)))];
-            tensor<fp16, [1, 1024, 1, 188]> value_43_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = value_43_dilations_0, groups = value_43_groups_0, pad = value_43_pad_0, pad_type = value_43_pad_type_0, strides = value_43_strides_0, weight = layers_21_self_attn_v_proj_weight_to_fp16, x = obj_87_cast_fp16)[name = tensor<string, []>("value_43_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_6162_to_fp16 = const()[name = tensor<string, []>("op_6162_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1090472384)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_87_cast_fp16 = add(x = query_85_cast_fp16, y = var_6162_to_fp16)[name = tensor<string, []>("query_87_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_6165_to_fp16 = const()[name = tensor<string, []>("op_6165_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1090474496)))];
-            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_43_cast_fp16 = add(x = query_85_cast_fp16, y = var_6165_to_fp16)[name = tensor<string, []>("q_with_bias_v_43_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> value_43_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_43_dilations_0, groups = value_43_groups_0, pad = value_43_pad_0, pad_type = value_43_pad_type_0, strides = value_43_strides_0, weight = layers_21_self_attn_v_proj_weight_to_fp16, x = obj_87_cast_fp16)[name = tensor<string, []>("value_43_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_6160_to_fp16 = const()[name = tensor<string, []>("op_6160_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1090472384)))];
+            tensor<fp16, [1, 1024, 1, 188]> query_87_cast_fp16 = add(x = query_85_cast_fp16, y = var_6160_to_fp16)[name = tensor<string, []>("query_87_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_6163_to_fp16 = const()[name = tensor<string, []>("op_6163_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1090474496)))];
+            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_43_cast_fp16 = add(x = query_85_cast_fp16, y = var_6163_to_fp16)[name = tensor<string, []>("q_with_bias_v_43_cast_fp16")];
             tensor<string, []> p_43_pad_type_0 = const()[name = tensor<string, []>("p_43_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> p_43_strides_0 = const()[name = tensor<string, []>("p_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> p_43_pad_0 = const()[name = tensor<string, []>("p_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> p_43_dilations_0 = const()[name = tensor<string, []>("p_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> p_43_groups_0 = const()[name = tensor<string, []>("p_43_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_21_self_attn_linear_pos_weight_to_fp16 = const()[name = tensor<string, []>("layers_21_self_attn_linear_pos_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1090476608)))];
-            tensor<fp16, [1, 1024, 1, 375]> p_43_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = p_43_dilations_0, groups = p_43_groups_0, pad = p_43_pad_0, pad_type = p_43_pad_type_0, strides = p_43_strides_0, weight = layers_21_self_attn_linear_pos_weight_to_fp16, x = pos_enc_to_fp16)[name = tensor<string, []>("p_43_cast_fp16")];
-            tensor<int32, [4]> var_6176 = const()[name = tensor<string, []>("op_6176"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_6177_cast_fp16 = reshape(shape = var_6176, x = q_with_bias_v_43_cast_fp16)[name = tensor<string, []>("op_6177_cast_fp16")];
-            tensor<int32, [4]> var_6178 = const()[name = tensor<string, []>("op_6178"), val = tensor<int32, [4]>([1, 8, 128, -1])];
-            tensor<fp16, [1, 8, 128, 375]> var_6179_cast_fp16 = reshape(shape = var_6178, x = p_43_cast_fp16)[name = tensor<string, []>("op_6179_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 375]> p_43_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_43_dilations_0, groups = p_43_groups_0, pad = p_43_pad_0, pad_type = p_43_pad_type_0, strides = p_43_strides_0, weight = layers_21_self_attn_linear_pos_weight_to_fp16, x = obj_3_cast_fp16)[name = tensor<string, []>("p_43_cast_fp16")];
+            tensor<int32, [4]> var_6174 = const()[name = tensor<string, []>("op_6174"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_6175_cast_fp16 = reshape(shape = var_6174, x = q_with_bias_v_43_cast_fp16)[name = tensor<string, []>("op_6175_cast_fp16")];
+            tensor<int32, [4]> var_6176 = const()[name = tensor<string, []>("op_6176"), val = tensor<int32, [4]>([1, 8, 128, -1])];
+            tensor<fp16, [1, 8, 128, 375]> var_6177_cast_fp16 = reshape(shape = var_6176, x = p_43_cast_fp16)[name = tensor<string, []>("op_6177_cast_fp16")];
             tensor<bool, []> matrix_bd_169_transpose_x_0 = const()[name = tensor<string, []>("matrix_bd_169_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> matrix_bd_169_transpose_y_0 = const()[name = tensor<string, []>("matrix_bd_169_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_169_cast_fp16 = matmul(transpose_x = matrix_bd_169_transpose_x_0, transpose_y = matrix_bd_169_transpose_y_0, x = var_6177_cast_fp16, y = var_6179_cast_fp16)[name = tensor<string, []>("matrix_bd_169_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_169_cast_fp16 = matmul(transpose_x = matrix_bd_169_transpose_x_0, transpose_y = matrix_bd_169_transpose_y_0, x = var_6175_cast_fp16, y = var_6177_cast_fp16)[name = tensor<string, []>("matrix_bd_169_cast_fp16")];
             tensor<int32, [8]> matrix_bd_171_pad_0 = const()[name = tensor<string, []>("matrix_bd_171_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
             tensor<string, []> matrix_bd_171_mode_0 = const()[name = tensor<string, []>("matrix_bd_171_mode_0"), val = tensor<string, []>("constant")];
             tensor<fp16, []> const_241_to_fp16 = const()[name = tensor<string, []>("const_241_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
             tensor<fp16, [1, 8, 188, 376]> matrix_bd_171_cast_fp16 = pad(constant_val = const_241_to_fp16, mode = matrix_bd_171_mode_0, pad = matrix_bd_171_pad_0, x = matrix_bd_169_cast_fp16)[name = tensor<string, []>("matrix_bd_171_cast_fp16")];
-            tensor<int32, [4]> var_6188 = const()[name = tensor<string, []>("op_6188"), val = tensor<int32, [4]>([1, 8, -1, 188])];
-            tensor<fp16, [1, 8, 376, 188]> matrix_bd_173_cast_fp16 = reshape(shape = var_6188, x = matrix_bd_171_cast_fp16)[name = tensor<string, []>("matrix_bd_173_cast_fp16")];
-            tensor<int32, [4]> var_6192_begin_0 = const()[name = tensor<string, []>("op_6192_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
-            tensor<int32, [4]> var_6192_end_0 = const()[name = tensor<string, []>("op_6192_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
-            tensor<bool, [4]> var_6192_end_mask_0 = const()[name = tensor<string, []>("op_6192_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
-            tensor<fp16, [1, 8, 375, 188]> var_6192_cast_fp16 = slice_by_index(begin = var_6192_begin_0, end = var_6192_end_0, end_mask = var_6192_end_mask_0, x = matrix_bd_173_cast_fp16)[name = tensor<string, []>("op_6192_cast_fp16")];
-            tensor<int32, [4]> var_6193 = const()[name = tensor<string, []>("op_6193"), val = tensor<int32, [4]>([1, 8, 188, 375])];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_175_cast_fp16 = reshape(shape = var_6193, x = var_6192_cast_fp16)[name = tensor<string, []>("matrix_bd_175_cast_fp16")];
-            tensor<int32, [4]> var_6198_begin_0 = const()[name = tensor<string, []>("op_6198_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [4]> var_6198_end_0 = const()[name = tensor<string, []>("op_6198_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
-            tensor<bool, [4]> var_6198_end_mask_0 = const()[name = tensor<string, []>("op_6198_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
-            tensor<fp16, [1, 8, 188, 188]> var_6198_cast_fp16 = slice_by_index(begin = var_6198_begin_0, end = var_6198_end_0, end_mask = var_6198_end_mask_0, x = matrix_bd_175_cast_fp16)[name = tensor<string, []>("op_6198_cast_fp16")];
-            tensor<fp16, []> var_6199_to_fp16 = const()[name = tensor<string, []>("op_6199_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 188, 188]> qk_mask_43_cast_fp16 = mul(x = var_6198_cast_fp16, y = var_6199_to_fp16)[name = tensor<string, []>("qk_mask_43_cast_fp16")];
-            tensor<int32, [4]> var_6203 = const()[name = tensor<string, []>("op_6203"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> mh_q_43_cast_fp16 = reshape(shape = var_6203, x = query_87_cast_fp16)[name = tensor<string, []>("mh_q_43_cast_fp16")];
-            tensor<fp16, []> var_6205_to_fp16 = const()[name = tensor<string, []>("op_6205_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 128, 188]> var_6206_cast_fp16 = mul(x = mh_q_43_cast_fp16, y = var_6205_to_fp16)[name = tensor<string, []>("op_6206_cast_fp16")];
-            tensor<int32, [4]> var_6209 = const()[name = tensor<string, []>("op_6209"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_6210_cast_fp16 = reshape(shape = var_6209, x = key_43_cast_fp16)[name = tensor<string, []>("op_6210_cast_fp16")];
+            tensor<int32, [4]> var_6186 = const()[name = tensor<string, []>("op_6186"), val = tensor<int32, [4]>([1, 8, -1, 188])];
+            tensor<fp16, [1, 8, 376, 188]> matrix_bd_173_cast_fp16 = reshape(shape = var_6186, x = matrix_bd_171_cast_fp16)[name = tensor<string, []>("matrix_bd_173_cast_fp16")];
+            tensor<int32, [4]> var_6190_begin_0 = const()[name = tensor<string, []>("op_6190_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_6190_end_0 = const()[name = tensor<string, []>("op_6190_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
+            tensor<bool, [4]> var_6190_end_mask_0 = const()[name = tensor<string, []>("op_6190_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 375, 188]> var_6190_cast_fp16 = slice_by_index(begin = var_6190_begin_0, end = var_6190_end_0, end_mask = var_6190_end_mask_0, x = matrix_bd_173_cast_fp16)[name = tensor<string, []>("op_6190_cast_fp16")];
+            tensor<int32, [4]> var_6191 = const()[name = tensor<string, []>("op_6191"), val = tensor<int32, [4]>([1, 8, 188, 375])];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_175_cast_fp16 = reshape(shape = var_6191, x = var_6190_cast_fp16)[name = tensor<string, []>("matrix_bd_175_cast_fp16")];
+            tensor<int32, [4]> var_6196_begin_0 = const()[name = tensor<string, []>("op_6196_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_6196_end_0 = const()[name = tensor<string, []>("op_6196_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
+            tensor<bool, [4]> var_6196_end_mask_0 = const()[name = tensor<string, []>("op_6196_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 188, 188]> var_6196_cast_fp16 = slice_by_index(begin = var_6196_begin_0, end = var_6196_end_0, end_mask = var_6196_end_mask_0, x = matrix_bd_175_cast_fp16)[name = tensor<string, []>("op_6196_cast_fp16")];
+            tensor<fp16, []> var_6197_to_fp16 = const()[name = tensor<string, []>("op_6197_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 188, 188]> qk_mask_43_cast_fp16 = mul(x = var_6196_cast_fp16, y = var_6197_to_fp16)[name = tensor<string, []>("qk_mask_43_cast_fp16")];
+            tensor<int32, [4]> var_6201 = const()[name = tensor<string, []>("op_6201"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> mh_q_43_cast_fp16 = reshape(shape = var_6201, x = query_87_cast_fp16)[name = tensor<string, []>("mh_q_43_cast_fp16")];
+            tensor<fp16, []> var_6203_to_fp16 = const()[name = tensor<string, []>("op_6203_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 128, 188]> var_6204_cast_fp16 = mul(x = mh_q_43_cast_fp16, y = var_6203_to_fp16)[name = tensor<string, []>("op_6204_cast_fp16")];
+            tensor<int32, [4]> var_6207 = const()[name = tensor<string, []>("op_6207"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_6208_cast_fp16 = reshape(shape = var_6207, x = key_43_cast_fp16)[name = tensor<string, []>("op_6208_cast_fp16")];
             tensor<bool, []> mh_w_85_transpose_x_0 = const()[name = tensor<string, []>("mh_w_85_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> mh_w_85_transpose_y_0 = const()[name = tensor<string, []>("mh_w_85_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 188]> mh_w_85_cast_fp16 = matmul(transpose_x = mh_w_85_transpose_x_0, transpose_y = mh_w_85_transpose_y_0, x = var_6206_cast_fp16, y = var_6210_cast_fp16)[name = tensor<string, []>("mh_w_85_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> mh_w_85_cast_fp16 = matmul(transpose_x = mh_w_85_transpose_x_0, transpose_y = mh_w_85_transpose_y_0, x = var_6204_cast_fp16, y = var_6208_cast_fp16)[name = tensor<string, []>("mh_w_85_cast_fp16")];
             tensor<fp16, [1, 8, 188, 188]> mh_w_87_cast_fp16 = add(x = mh_w_85_cast_fp16, y = qk_mask_43_cast_fp16)[name = tensor<string, []>("mh_w_87_cast_fp16")];
-            tensor<fp16, [1, 8, 188, 188]> var_6214_cast_fp16 = softmax(axis = var_6055, x = mh_w_87_cast_fp16)[name = tensor<string, []>("op_6214_cast_fp16")];
-            tensor<int32, [4]> var_6215 = const()[name = tensor<string, []>("op_6215"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_6216_cast_fp16 = reshape(shape = var_6215, x = value_43_cast_fp16)[name = tensor<string, []>("op_6216_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> var_6212_cast_fp16 = softmax(axis = var_6053, x = mh_w_87_cast_fp16)[name = tensor<string, []>("op_6212_cast_fp16")];
+            tensor<int32, [4]> var_6213 = const()[name = tensor<string, []>("op_6213"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_6214_cast_fp16 = reshape(shape = var_6213, x = value_43_cast_fp16)[name = tensor<string, []>("op_6214_cast_fp16")];
             tensor<bool, []> attn_43_transpose_x_0 = const()[name = tensor<string, []>("attn_43_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> attn_43_transpose_y_0 = const()[name = tensor<string, []>("attn_43_transpose_y_0"), val = tensor<bool, []>(true)];
-            tensor<fp16, [1, 8, 128, 188]> attn_43_cast_fp16 = matmul(transpose_x = attn_43_transpose_x_0, transpose_y = attn_43_transpose_y_0, x = var_6216_cast_fp16, y = var_6214_cast_fp16)[name = tensor<string, []>("attn_43_cast_fp16")];
-            tensor<int32, [4]> var_6219 = const()[name = tensor<string, []>("op_6219"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
-            tensor<fp16, [1, 1024, 1, 188]> input_567_cast_fp16 = reshape(shape = var_6219, x = attn_43_cast_fp16)[name = tensor<string, []>("input_567_cast_fp16")];
+            tensor<fp16, [1, 8, 128, 188]> attn_43_cast_fp16 = matmul(transpose_x = attn_43_transpose_x_0, transpose_y = attn_43_transpose_y_0, x = var_6214_cast_fp16, y = var_6212_cast_fp16)[name = tensor<string, []>("attn_43_cast_fp16")];
+            tensor<int32, [4]> var_6217 = const()[name = tensor<string, []>("op_6217"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
+            tensor<fp16, [1, 1024, 1, 188]> input_569_cast_fp16 = reshape(shape = var_6217, x = attn_43_cast_fp16)[name = tensor<string, []>("input_569_cast_fp16")];
             tensor<string, []> obj_89_pad_type_0 = const()[name = tensor<string, []>("obj_89_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> obj_89_strides_0 = const()[name = tensor<string, []>("obj_89_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> obj_89_pad_0 = const()[name = tensor<string, []>("obj_89_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> obj_89_dilations_0 = const()[name = tensor<string, []>("obj_89_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> obj_89_groups_0 = const()[name = tensor<string, []>("obj_89_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_21_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_21_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1092573824)))];
-            tensor<fp16, [1, 1024, 1, 188]> obj_89_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = obj_89_dilations_0, groups = obj_89_groups_0, pad = obj_89_pad_0, pad_type = obj_89_pad_type_0, strides = obj_89_strides_0, weight = layers_21_self_attn_o_proj_weight_to_fp16, x = input_567_cast_fp16)[name = tensor<string, []>("obj_89_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_89_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_89_dilations_0, groups = obj_89_groups_0, pad = obj_89_pad_0, pad_type = obj_89_pad_type_0, strides = obj_89_strides_0, weight = layers_21_self_attn_o_proj_weight_to_fp16, x = input_569_cast_fp16)[name = tensor<string, []>("obj_89_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_215_cast_fp16 = add(x = inputs_213_cast_fp16, y = obj_89_cast_fp16)[name = tensor<string, []>("inputs_215_cast_fp16")];
             tensor<int32, [1]> out_215_axes_0 = const()[name = tensor<string, []>("out_215_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_6237_to_fp16 = const()[name = tensor<string, []>("op_6237_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_215_cast_fp16 = layer_norm(axes = out_215_axes_0, epsilon = var_6237_to_fp16, x = inputs_215_cast_fp16)[name = tensor<string, []>("out_215_cast_fp16")];
-            tensor<fp16, [1024]> input_569_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_569_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1094671040)))];
-            tensor<fp16, [1024]> input_569_beta_0_to_fp16 = const()[name = tensor<string, []>("input_569_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1094673152)))];
-            tensor<fp16, []> input_569_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_569_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_569_cast_fp16 = batch_norm(beta = input_569_beta_0_to_fp16, epsilon = input_569_epsilon_0_to_fp16, gamma = input_569_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_215_cast_fp16)[name = tensor<string, []>("input_569_cast_fp16")];
-            tensor<string, []> input_571_pad_type_0 = const()[name = tensor<string, []>("input_571_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_571_strides_0 = const()[name = tensor<string, []>("input_571_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_571_pad_0 = const()[name = tensor<string, []>("input_571_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_571_dilations_0 = const()[name = tensor<string, []>("input_571_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_571_groups_0 = const()[name = tensor<string, []>("input_571_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_6235_to_fp16 = const()[name = tensor<string, []>("op_6235_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_215_cast_fp16 = layer_norm(axes = out_215_axes_0, epsilon = var_6235_to_fp16, x = inputs_215_cast_fp16)[name = tensor<string, []>("out_215_cast_fp16")];
+            tensor<fp16, [1024]> input_571_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_571_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1094671040)))];
+            tensor<fp16, [1024]> input_571_beta_0_to_fp16 = const()[name = tensor<string, []>("input_571_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1094673152)))];
+            tensor<fp16, []> input_571_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_571_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_571_cast_fp16 = batch_norm(beta = input_571_beta_0_to_fp16, epsilon = input_571_epsilon_0_to_fp16, gamma = input_571_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_215_cast_fp16)[name = tensor<string, []>("input_571_cast_fp16")];
+            tensor<string, []> input_573_pad_type_0 = const()[name = tensor<string, []>("input_573_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_573_strides_0 = const()[name = tensor<string, []>("input_573_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_573_pad_0 = const()[name = tensor<string, []>("input_573_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_573_dilations_0 = const()[name = tensor<string, []>("input_573_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_573_groups_0 = const()[name = tensor<string, []>("input_573_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [2048, 1024, 1, 1]> layers_21_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("layers_21_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1094675264)))];
-            tensor<fp16, [1, 2048, 1, 188]> input_571_cast_fp16 = conv(dilations = input_571_dilations_0, groups = input_571_groups_0, pad = input_571_pad_0, pad_type = input_571_pad_type_0, strides = input_571_strides_0, weight = layers_21_conv_pointwise_conv1_weight_to_fp16, x = input_569_cast_fp16)[name = tensor<string, []>("input_571_cast_fp16")];
-            tensor<int32, []> input_573_split_num_splits_0 = const()[name = tensor<string, []>("input_573_split_num_splits_0"), val = tensor<int32, []>(2)];
-            tensor<int32, []> input_573_split_axis_0 = const()[name = tensor<string, []>("input_573_split_axis_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1, 1024, 1, 188]> input_573_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_573_split_cast_fp16_1 = split(axis = input_573_split_axis_0, num_splits = input_573_split_num_splits_0, x = input_571_cast_fp16)[name = tensor<string, []>("input_573_split_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_573_split_1_sigmoid_cast_fp16 = sigmoid(x = input_573_split_cast_fp16_1)[name = tensor<string, []>("input_573_split_1_sigmoid_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_573_cast_fp16 = mul(x = input_573_split_cast_fp16_0, y = input_573_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_573_cast_fp16")];
-            tensor<string, []> input_575_pad_type_0 = const()[name = tensor<string, []>("input_575_pad_type_0"), val = tensor<string, []>("custom")];
-            tensor<int32, [4]> input_575_pad_0 = const()[name = tensor<string, []>("input_575_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
-            tensor<int32, []> input_575_groups_0 = const()[name = tensor<string, []>("input_575_groups_0"), val = tensor<int32, []>(1024)];
-            tensor<int32, [2]> input_575_strides_0 = const()[name = tensor<string, []>("input_575_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> input_575_dilations_0 = const()[name = tensor<string, []>("input_575_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [1, 2048, 1, 188]> input_573_cast_fp16 = conv(dilations = input_573_dilations_0, groups = input_573_groups_0, pad = input_573_pad_0, pad_type = input_573_pad_type_0, strides = input_573_strides_0, weight = layers_21_conv_pointwise_conv1_weight_to_fp16, x = input_571_cast_fp16)[name = tensor<string, []>("input_573_cast_fp16")];
+            tensor<int32, []> input_575_split_num_splits_0 = const()[name = tensor<string, []>("input_575_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> input_575_split_axis_0 = const()[name = tensor<string, []>("input_575_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 1, 188]> input_575_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_575_split_cast_fp16_1 = split(axis = input_575_split_axis_0, num_splits = input_575_split_num_splits_0, x = input_573_cast_fp16)[name = tensor<string, []>("input_575_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_575_split_1_sigmoid_cast_fp16 = sigmoid(x = input_575_split_cast_fp16_1)[name = tensor<string, []>("input_575_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_575_cast_fp16 = mul(x = input_575_split_cast_fp16_0, y = input_575_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_575_cast_fp16")];
+            tensor<string, []> input_577_pad_type_0 = const()[name = tensor<string, []>("input_577_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_577_pad_0 = const()[name = tensor<string, []>("input_577_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
+            tensor<int32, []> input_577_groups_0 = const()[name = tensor<string, []>("input_577_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [2]> input_577_strides_0 = const()[name = tensor<string, []>("input_577_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> input_577_dilations_0 = const()[name = tensor<string, []>("input_577_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<fp16, [1024, 1, 1, 9]> const_310_to_fp16 = const()[name = tensor<string, []>("const_310_to_fp16"), val = tensor<fp16, [1024, 1, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1098869632)))];
             tensor<fp16, [1024]> const_311_to_fp16 = const()[name = tensor<string, []>("const_311_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1098888128)))];
-            tensor<fp16, [1, 1024, 1, 188]> input_577_cast_fp16 = conv(bias = const_311_to_fp16, dilations = input_575_dilations_0, groups = input_575_groups_0, pad = input_575_pad_0, pad_type = input_575_pad_type_0, strides = input_575_strides_0, weight = const_310_to_fp16, x = input_573_cast_fp16)[name = tensor<string, []>("input_577_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_579_cast_fp16 = silu(x = input_577_cast_fp16)[name = tensor<string, []>("input_579_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_579_cast_fp16 = conv(bias = const_311_to_fp16, dilations = input_577_dilations_0, groups = input_577_groups_0, pad = input_577_pad_0, pad_type = input_577_pad_type_0, strides = input_577_strides_0, weight = const_310_to_fp16, x = input_575_cast_fp16)[name = tensor<string, []>("input_579_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_581_cast_fp16 = silu(x = input_579_cast_fp16)[name = tensor<string, []>("input_581_cast_fp16")];
             tensor<string, []> x_131_pad_type_0 = const()[name = tensor<string, []>("x_131_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_131_strides_0 = const()[name = tensor<string, []>("x_131_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_131_pad_0 = const()[name = tensor<string, []>("x_131_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_131_dilations_0 = const()[name = tensor<string, []>("x_131_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_131_groups_0 = const()[name = tensor<string, []>("x_131_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_21_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("layers_21_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1098890240)))];
-            tensor<fp16, [1, 1024, 1, 188]> x_131_cast_fp16 = conv(dilations = x_131_dilations_0, groups = x_131_groups_0, pad = x_131_pad_0, pad_type = x_131_pad_type_0, strides = x_131_strides_0, weight = layers_21_conv_pointwise_conv2_weight_to_fp16, x = input_579_cast_fp16)[name = tensor<string, []>("x_131_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> x_131_cast_fp16 = conv(dilations = x_131_dilations_0, groups = x_131_groups_0, pad = x_131_pad_0, pad_type = x_131_pad_type_0, strides = x_131_strides_0, weight = layers_21_conv_pointwise_conv2_weight_to_fp16, x = input_581_cast_fp16)[name = tensor<string, []>("x_131_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_217_cast_fp16 = add(x = inputs_215_cast_fp16, y = x_131_cast_fp16)[name = tensor<string, []>("inputs_217_cast_fp16")];
             tensor<int32, [1]> out_217_axes_0 = const()[name = tensor<string, []>("out_217_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_6285_to_fp16 = const()[name = tensor<string, []>("op_6285_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_217_cast_fp16 = layer_norm(axes = out_217_axes_0, epsilon = var_6285_to_fp16, x = inputs_217_cast_fp16)[name = tensor<string, []>("out_217_cast_fp16")];
-            tensor<fp16, [1024]> input_581_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_581_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1100987456)))];
-            tensor<fp16, [1024]> input_581_beta_0_to_fp16 = const()[name = tensor<string, []>("input_581_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1100989568)))];
-            tensor<fp16, []> input_581_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_581_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_581_cast_fp16 = batch_norm(beta = input_581_beta_0_to_fp16, epsilon = input_581_epsilon_0_to_fp16, gamma = input_581_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_217_cast_fp16)[name = tensor<string, []>("input_581_cast_fp16")];
-            tensor<string, []> input_583_pad_type_0 = const()[name = tensor<string, []>("input_583_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_583_strides_0 = const()[name = tensor<string, []>("input_583_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_583_pad_0 = const()[name = tensor<string, []>("input_583_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_583_dilations_0 = const()[name = tensor<string, []>("input_583_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_583_groups_0 = const()[name = tensor<string, []>("input_583_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_6283_to_fp16 = const()[name = tensor<string, []>("op_6283_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_217_cast_fp16 = layer_norm(axes = out_217_axes_0, epsilon = var_6283_to_fp16, x = inputs_217_cast_fp16)[name = tensor<string, []>("out_217_cast_fp16")];
+            tensor<fp16, [1024]> input_583_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_583_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1100987456)))];
+            tensor<fp16, [1024]> input_583_beta_0_to_fp16 = const()[name = tensor<string, []>("input_583_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1100989568)))];
+            tensor<fp16, []> input_583_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_583_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_583_cast_fp16 = batch_norm(beta = input_583_beta_0_to_fp16, epsilon = input_583_epsilon_0_to_fp16, gamma = input_583_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_217_cast_fp16)[name = tensor<string, []>("input_583_cast_fp16")];
+            tensor<string, []> input_585_pad_type_0 = const()[name = tensor<string, []>("input_585_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_585_strides_0 = const()[name = tensor<string, []>("input_585_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_585_pad_0 = const()[name = tensor<string, []>("input_585_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_585_dilations_0 = const()[name = tensor<string, []>("input_585_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_585_groups_0 = const()[name = tensor<string, []>("input_585_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_21_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_21_feed_forward2_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1100991680)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_583_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_583_dilations_0, groups = input_583_groups_0, pad = input_583_pad_0, pad_type = input_583_pad_type_0, strides = input_583_strides_0, weight = layers_21_feed_forward2_fc1_weight_to_fp16, x = input_581_cast_fp16)[name = tensor<string, []>("input_583_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_585_cast_fp16 = silu(x = input_583_cast_fp16)[name = tensor<string, []>("input_585_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_585_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_585_dilations_0, groups = input_585_groups_0, pad = input_585_pad_0, pad_type = input_585_pad_type_0, strides = input_585_strides_0, weight = layers_21_feed_forward2_fc1_weight_to_fp16, x = input_583_cast_fp16)[name = tensor<string, []>("input_585_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_587_cast_fp16 = silu(x = input_585_cast_fp16)[name = tensor<string, []>("input_587_cast_fp16")];
             tensor<string, []> x_133_pad_type_0 = const()[name = tensor<string, []>("x_133_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_133_strides_0 = const()[name = tensor<string, []>("x_133_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_133_pad_0 = const()[name = tensor<string, []>("x_133_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_133_dilations_0 = const()[name = tensor<string, []>("x_133_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_133_groups_0 = const()[name = tensor<string, []>("x_133_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_6313_weight_0_to_fp16 = const()[name = tensor<string, []>("op_6313_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1109380352)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_6313_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_133_dilations_0, groups = x_133_groups_0, pad = x_133_pad_0, pad_type = x_133_pad_type_0, strides = x_133_strides_0, weight = var_6313_weight_0_to_fp16, x = input_585_cast_fp16)[name = tensor<string, []>("op_6313_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_219_cast_fp16 = add(x = inputs_217_cast_fp16, y = var_6313_cast_fp16)[name = tensor<string, []>("inputs_219_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_6311_weight_0_to_fp16 = const()[name = tensor<string, []>("op_6311_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1109380352)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_6311_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_133_dilations_0, groups = x_133_groups_0, pad = x_133_pad_0, pad_type = x_133_pad_type_0, strides = x_133_strides_0, weight = var_6311_weight_0_to_fp16, x = input_587_cast_fp16)[name = tensor<string, []>("op_6311_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_219_cast_fp16 = add(x = inputs_217_cast_fp16, y = var_6311_cast_fp16)[name = tensor<string, []>("inputs_219_cast_fp16")];
             tensor<int32, [1]> out_219_axes_0 = const()[name = tensor<string, []>("out_219_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_6323_to_fp16 = const()[name = tensor<string, []>("op_6323_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_219_cast_fp16 = layer_norm(axes = out_219_axes_0, epsilon = var_6323_to_fp16, x = inputs_219_cast_fp16)[name = tensor<string, []>("out_219_cast_fp16")];
+            tensor<fp16, []> var_6321_to_fp16 = const()[name = tensor<string, []>("op_6321_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_219_cast_fp16 = layer_norm(axes = out_219_axes_0, epsilon = var_6321_to_fp16, x = inputs_219_cast_fp16)[name = tensor<string, []>("out_219_cast_fp16")];
             tensor<fp16, [1024]> inputs_221_gamma_0_to_fp16 = const()[name = tensor<string, []>("inputs_221_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1117769024)))];
             tensor<fp16, [1024]> inputs_221_beta_0_to_fp16 = const()[name = tensor<string, []>("inputs_221_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1117771136)))];
             tensor<fp16, []> inputs_221_epsilon_0_to_fp16 = const()[name = tensor<string, []>("inputs_221_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_221_cast_fp16 = batch_norm(beta = inputs_221_beta_0_to_fp16, epsilon = inputs_221_epsilon_0_to_fp16, gamma = inputs_221_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_219_cast_fp16)[name = tensor<string, []>("inputs_221_cast_fp16")];
-            tensor<int32, []> var_6337 = const()[name = tensor<string, []>("op_6337"), val = tensor<int32, []>(3)];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_221_cast_fp16 = batch_norm(beta = inputs_221_beta_0_to_fp16, epsilon = inputs_221_epsilon_0_to_fp16, gamma = inputs_221_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_219_cast_fp16)[name = tensor<string, []>("inputs_221_cast_fp16")];
+            tensor<int32, []> var_6335 = const()[name = tensor<string, []>("op_6335"), val = tensor<int32, []>(3)];
             tensor<int32, [1]> out_221_axes_0 = const()[name = tensor<string, []>("out_221_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_6368_to_fp16 = const()[name = tensor<string, []>("op_6368_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_221_cast_fp16 = layer_norm(axes = out_221_axes_0, epsilon = var_6368_to_fp16, x = inputs_221_cast_fp16)[name = tensor<string, []>("out_221_cast_fp16")];
-            tensor<fp16, [1024]> input_587_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_587_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1117773248)))];
-            tensor<fp16, [1024]> input_587_beta_0_to_fp16 = const()[name = tensor<string, []>("input_587_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1117775360)))];
-            tensor<fp16, []> input_587_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_587_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_587_cast_fp16 = batch_norm(beta = input_587_beta_0_to_fp16, epsilon = input_587_epsilon_0_to_fp16, gamma = input_587_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_221_cast_fp16)[name = tensor<string, []>("input_587_cast_fp16")];
-            tensor<string, []> input_589_pad_type_0 = const()[name = tensor<string, []>("input_589_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_589_strides_0 = const()[name = tensor<string, []>("input_589_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_589_pad_0 = const()[name = tensor<string, []>("input_589_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_589_dilations_0 = const()[name = tensor<string, []>("input_589_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_589_groups_0 = const()[name = tensor<string, []>("input_589_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_6366_to_fp16 = const()[name = tensor<string, []>("op_6366_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_221_cast_fp16 = layer_norm(axes = out_221_axes_0, epsilon = var_6366_to_fp16, x = inputs_221_cast_fp16)[name = tensor<string, []>("out_221_cast_fp16")];
+            tensor<fp16, [1024]> input_589_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_589_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1117773248)))];
+            tensor<fp16, [1024]> input_589_beta_0_to_fp16 = const()[name = tensor<string, []>("input_589_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1117775360)))];
+            tensor<fp16, []> input_589_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_589_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_589_cast_fp16 = batch_norm(beta = input_589_beta_0_to_fp16, epsilon = input_589_epsilon_0_to_fp16, gamma = input_589_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_221_cast_fp16)[name = tensor<string, []>("input_589_cast_fp16")];
+            tensor<string, []> input_591_pad_type_0 = const()[name = tensor<string, []>("input_591_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_591_strides_0 = const()[name = tensor<string, []>("input_591_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_591_pad_0 = const()[name = tensor<string, []>("input_591_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_591_dilations_0 = const()[name = tensor<string, []>("input_591_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_591_groups_0 = const()[name = tensor<string, []>("input_591_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_22_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_22_feed_forward1_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1117777472)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_589_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_589_dilations_0, groups = input_589_groups_0, pad = input_589_pad_0, pad_type = input_589_pad_type_0, strides = input_589_strides_0, weight = layers_22_feed_forward1_fc1_weight_to_fp16, x = input_587_cast_fp16)[name = tensor<string, []>("input_589_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_591_cast_fp16 = silu(x = input_589_cast_fp16)[name = tensor<string, []>("input_591_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_591_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_591_dilations_0, groups = input_591_groups_0, pad = input_591_pad_0, pad_type = input_591_pad_type_0, strides = input_591_strides_0, weight = layers_22_feed_forward1_fc1_weight_to_fp16, x = input_589_cast_fp16)[name = tensor<string, []>("input_591_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_593_cast_fp16 = silu(x = input_591_cast_fp16)[name = tensor<string, []>("input_593_cast_fp16")];
             tensor<string, []> x_135_pad_type_0 = const()[name = tensor<string, []>("x_135_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_135_strides_0 = const()[name = tensor<string, []>("x_135_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_135_pad_0 = const()[name = tensor<string, []>("x_135_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_135_dilations_0 = const()[name = tensor<string, []>("x_135_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_135_groups_0 = const()[name = tensor<string, []>("x_135_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_6396_weight_0_to_fp16 = const()[name = tensor<string, []>("op_6396_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1126166144)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_6396_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_135_dilations_0, groups = x_135_groups_0, pad = x_135_pad_0, pad_type = x_135_pad_type_0, strides = x_135_strides_0, weight = var_6396_weight_0_to_fp16, x = input_591_cast_fp16)[name = tensor<string, []>("op_6396_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_223_cast_fp16 = add(x = inputs_221_cast_fp16, y = var_6396_cast_fp16)[name = tensor<string, []>("inputs_223_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_6394_weight_0_to_fp16 = const()[name = tensor<string, []>("op_6394_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1126166144)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_6394_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_135_dilations_0, groups = x_135_groups_0, pad = x_135_pad_0, pad_type = x_135_pad_type_0, strides = x_135_strides_0, weight = var_6394_weight_0_to_fp16, x = input_593_cast_fp16)[name = tensor<string, []>("op_6394_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_223_cast_fp16 = add(x = inputs_221_cast_fp16, y = var_6394_cast_fp16)[name = tensor<string, []>("inputs_223_cast_fp16")];
             tensor<int32, [1]> out_223_axes_0 = const()[name = tensor<string, []>("out_223_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_6406_to_fp16 = const()[name = tensor<string, []>("op_6406_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_223_cast_fp16 = layer_norm(axes = out_223_axes_0, epsilon = var_6406_to_fp16, x = inputs_223_cast_fp16)[name = tensor<string, []>("out_223_cast_fp16")];
+            tensor<fp16, []> var_6404_to_fp16 = const()[name = tensor<string, []>("op_6404_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_223_cast_fp16 = layer_norm(axes = out_223_axes_0, epsilon = var_6404_to_fp16, x = inputs_223_cast_fp16)[name = tensor<string, []>("out_223_cast_fp16")];
             tensor<fp16, [1024]> obj_91_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_91_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1134554816)))];
             tensor<fp16, [1024]> obj_91_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_91_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1134556928)))];
             tensor<fp16, []> obj_91_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_91_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> obj_91_cast_fp16 = batch_norm(beta = obj_91_beta_0_to_fp16, epsilon = obj_91_epsilon_0_to_fp16, gamma = obj_91_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_223_cast_fp16)[name = tensor<string, []>("obj_91_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_91_cast_fp16 = batch_norm(beta = obj_91_beta_0_to_fp16, epsilon = obj_91_epsilon_0_to_fp16, gamma = obj_91_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_223_cast_fp16)[name = tensor<string, []>("obj_91_cast_fp16")];
             tensor<string, []> query_89_pad_type_0 = const()[name = tensor<string, []>("query_89_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> query_89_strides_0 = const()[name = tensor<string, []>("query_89_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> query_89_pad_0 = const()[name = tensor<string, []>("query_89_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> query_89_dilations_0 = const()[name = tensor<string, []>("query_89_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> query_89_groups_0 = const()[name = tensor<string, []>("query_89_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_22_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_22_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1134559040)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_89_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = query_89_dilations_0, groups = query_89_groups_0, pad = query_89_pad_0, pad_type = query_89_pad_type_0, strides = query_89_strides_0, weight = layers_22_self_attn_q_proj_weight_to_fp16, x = obj_91_cast_fp16)[name = tensor<string, []>("query_89_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> query_89_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_89_dilations_0, groups = query_89_groups_0, pad = query_89_pad_0, pad_type = query_89_pad_type_0, strides = query_89_strides_0, weight = layers_22_self_attn_q_proj_weight_to_fp16, x = obj_91_cast_fp16)[name = tensor<string, []>("query_89_cast_fp16")];
             tensor<string, []> key_45_pad_type_0 = const()[name = tensor<string, []>("key_45_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> key_45_strides_0 = const()[name = tensor<string, []>("key_45_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> key_45_pad_0 = const()[name = tensor<string, []>("key_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
@@ -4072,173 +4074,173 @@ program(1.0)
             tensor<int32, [2]> value_45_dilations_0 = const()[name = tensor<string, []>("value_45_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> value_45_groups_0 = const()[name = tensor<string, []>("value_45_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_22_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_22_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1138753472)))];
-            tensor<fp16, [1, 1024, 1, 188]> value_45_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = value_45_dilations_0, groups = value_45_groups_0, pad = value_45_pad_0, pad_type = value_45_pad_type_0, strides = value_45_strides_0, weight = layers_22_self_attn_v_proj_weight_to_fp16, x = obj_91_cast_fp16)[name = tensor<string, []>("value_45_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_6444_to_fp16 = const()[name = tensor<string, []>("op_6444_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1140850688)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_91_cast_fp16 = add(x = query_89_cast_fp16, y = var_6444_to_fp16)[name = tensor<string, []>("query_91_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_6447_to_fp16 = const()[name = tensor<string, []>("op_6447_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1140852800)))];
-            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_45_cast_fp16 = add(x = query_89_cast_fp16, y = var_6447_to_fp16)[name = tensor<string, []>("q_with_bias_v_45_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> value_45_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_45_dilations_0, groups = value_45_groups_0, pad = value_45_pad_0, pad_type = value_45_pad_type_0, strides = value_45_strides_0, weight = layers_22_self_attn_v_proj_weight_to_fp16, x = obj_91_cast_fp16)[name = tensor<string, []>("value_45_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_6442_to_fp16 = const()[name = tensor<string, []>("op_6442_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1140850688)))];
+            tensor<fp16, [1, 1024, 1, 188]> query_91_cast_fp16 = add(x = query_89_cast_fp16, y = var_6442_to_fp16)[name = tensor<string, []>("query_91_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_6445_to_fp16 = const()[name = tensor<string, []>("op_6445_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1140852800)))];
+            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_45_cast_fp16 = add(x = query_89_cast_fp16, y = var_6445_to_fp16)[name = tensor<string, []>("q_with_bias_v_45_cast_fp16")];
             tensor<string, []> p_45_pad_type_0 = const()[name = tensor<string, []>("p_45_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> p_45_strides_0 = const()[name = tensor<string, []>("p_45_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> p_45_pad_0 = const()[name = tensor<string, []>("p_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> p_45_dilations_0 = const()[name = tensor<string, []>("p_45_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> p_45_groups_0 = const()[name = tensor<string, []>("p_45_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_22_self_attn_linear_pos_weight_to_fp16 = const()[name = tensor<string, []>("layers_22_self_attn_linear_pos_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1140854912)))];
-            tensor<fp16, [1, 1024, 1, 375]> p_45_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = p_45_dilations_0, groups = p_45_groups_0, pad = p_45_pad_0, pad_type = p_45_pad_type_0, strides = p_45_strides_0, weight = layers_22_self_attn_linear_pos_weight_to_fp16, x = pos_enc_to_fp16)[name = tensor<string, []>("p_45_cast_fp16")];
-            tensor<int32, [4]> var_6458 = const()[name = tensor<string, []>("op_6458"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_6459_cast_fp16 = reshape(shape = var_6458, x = q_with_bias_v_45_cast_fp16)[name = tensor<string, []>("op_6459_cast_fp16")];
-            tensor<int32, [4]> var_6460 = const()[name = tensor<string, []>("op_6460"), val = tensor<int32, [4]>([1, 8, 128, -1])];
-            tensor<fp16, [1, 8, 128, 375]> var_6461_cast_fp16 = reshape(shape = var_6460, x = p_45_cast_fp16)[name = tensor<string, []>("op_6461_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 375]> p_45_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_45_dilations_0, groups = p_45_groups_0, pad = p_45_pad_0, pad_type = p_45_pad_type_0, strides = p_45_strides_0, weight = layers_22_self_attn_linear_pos_weight_to_fp16, x = obj_3_cast_fp16)[name = tensor<string, []>("p_45_cast_fp16")];
+            tensor<int32, [4]> var_6456 = const()[name = tensor<string, []>("op_6456"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_6457_cast_fp16 = reshape(shape = var_6456, x = q_with_bias_v_45_cast_fp16)[name = tensor<string, []>("op_6457_cast_fp16")];
+            tensor<int32, [4]> var_6458 = const()[name = tensor<string, []>("op_6458"), val = tensor<int32, [4]>([1, 8, 128, -1])];
+            tensor<fp16, [1, 8, 128, 375]> var_6459_cast_fp16 = reshape(shape = var_6458, x = p_45_cast_fp16)[name = tensor<string, []>("op_6459_cast_fp16")];
             tensor<bool, []> matrix_bd_177_transpose_x_0 = const()[name = tensor<string, []>("matrix_bd_177_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> matrix_bd_177_transpose_y_0 = const()[name = tensor<string, []>("matrix_bd_177_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_177_cast_fp16 = matmul(transpose_x = matrix_bd_177_transpose_x_0, transpose_y = matrix_bd_177_transpose_y_0, x = var_6459_cast_fp16, y = var_6461_cast_fp16)[name = tensor<string, []>("matrix_bd_177_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_177_cast_fp16 = matmul(transpose_x = matrix_bd_177_transpose_x_0, transpose_y = matrix_bd_177_transpose_y_0, x = var_6457_cast_fp16, y = var_6459_cast_fp16)[name = tensor<string, []>("matrix_bd_177_cast_fp16")];
             tensor<int32, [8]> matrix_bd_179_pad_0 = const()[name = tensor<string, []>("matrix_bd_179_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
             tensor<string, []> matrix_bd_179_mode_0 = const()[name = tensor<string, []>("matrix_bd_179_mode_0"), val = tensor<string, []>("constant")];
             tensor<fp16, []> const_252_to_fp16 = const()[name = tensor<string, []>("const_252_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
             tensor<fp16, [1, 8, 188, 376]> matrix_bd_179_cast_fp16 = pad(constant_val = const_252_to_fp16, mode = matrix_bd_179_mode_0, pad = matrix_bd_179_pad_0, x = matrix_bd_177_cast_fp16)[name = tensor<string, []>("matrix_bd_179_cast_fp16")];
-            tensor<int32, [4]> var_6470 = const()[name = tensor<string, []>("op_6470"), val = tensor<int32, [4]>([1, 8, -1, 188])];
-            tensor<fp16, [1, 8, 376, 188]> matrix_bd_181_cast_fp16 = reshape(shape = var_6470, x = matrix_bd_179_cast_fp16)[name = tensor<string, []>("matrix_bd_181_cast_fp16")];
-            tensor<int32, [4]> var_6474_begin_0 = const()[name = tensor<string, []>("op_6474_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
-            tensor<int32, [4]> var_6474_end_0 = const()[name = tensor<string, []>("op_6474_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
-            tensor<bool, [4]> var_6474_end_mask_0 = const()[name = tensor<string, []>("op_6474_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
-            tensor<fp16, [1, 8, 375, 188]> var_6474_cast_fp16 = slice_by_index(begin = var_6474_begin_0, end = var_6474_end_0, end_mask = var_6474_end_mask_0, x = matrix_bd_181_cast_fp16)[name = tensor<string, []>("op_6474_cast_fp16")];
-            tensor<int32, [4]> var_6475 = const()[name = tensor<string, []>("op_6475"), val = tensor<int32, [4]>([1, 8, 188, 375])];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_183_cast_fp16 = reshape(shape = var_6475, x = var_6474_cast_fp16)[name = tensor<string, []>("matrix_bd_183_cast_fp16")];
-            tensor<int32, [4]> var_6480_begin_0 = const()[name = tensor<string, []>("op_6480_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [4]> var_6480_end_0 = const()[name = tensor<string, []>("op_6480_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
-            tensor<bool, [4]> var_6480_end_mask_0 = const()[name = tensor<string, []>("op_6480_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
-            tensor<fp16, [1, 8, 188, 188]> var_6480_cast_fp16 = slice_by_index(begin = var_6480_begin_0, end = var_6480_end_0, end_mask = var_6480_end_mask_0, x = matrix_bd_183_cast_fp16)[name = tensor<string, []>("op_6480_cast_fp16")];
-            tensor<fp16, []> var_6481_to_fp16 = const()[name = tensor<string, []>("op_6481_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 188, 188]> qk_mask_45_cast_fp16 = mul(x = var_6480_cast_fp16, y = var_6481_to_fp16)[name = tensor<string, []>("qk_mask_45_cast_fp16")];
-            tensor<int32, [4]> var_6485 = const()[name = tensor<string, []>("op_6485"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> mh_q_45_cast_fp16 = reshape(shape = var_6485, x = query_91_cast_fp16)[name = tensor<string, []>("mh_q_45_cast_fp16")];
-            tensor<fp16, []> var_6487_to_fp16 = const()[name = tensor<string, []>("op_6487_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 128, 188]> var_6488_cast_fp16 = mul(x = mh_q_45_cast_fp16, y = var_6487_to_fp16)[name = tensor<string, []>("op_6488_cast_fp16")];
-            tensor<int32, [4]> var_6491 = const()[name = tensor<string, []>("op_6491"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_6492_cast_fp16 = reshape(shape = var_6491, x = key_45_cast_fp16)[name = tensor<string, []>("op_6492_cast_fp16")];
+            tensor<int32, [4]> var_6468 = const()[name = tensor<string, []>("op_6468"), val = tensor<int32, [4]>([1, 8, -1, 188])];
+            tensor<fp16, [1, 8, 376, 188]> matrix_bd_181_cast_fp16 = reshape(shape = var_6468, x = matrix_bd_179_cast_fp16)[name = tensor<string, []>("matrix_bd_181_cast_fp16")];
+            tensor<int32, [4]> var_6472_begin_0 = const()[name = tensor<string, []>("op_6472_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_6472_end_0 = const()[name = tensor<string, []>("op_6472_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
+            tensor<bool, [4]> var_6472_end_mask_0 = const()[name = tensor<string, []>("op_6472_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 375, 188]> var_6472_cast_fp16 = slice_by_index(begin = var_6472_begin_0, end = var_6472_end_0, end_mask = var_6472_end_mask_0, x = matrix_bd_181_cast_fp16)[name = tensor<string, []>("op_6472_cast_fp16")];
+            tensor<int32, [4]> var_6473 = const()[name = tensor<string, []>("op_6473"), val = tensor<int32, [4]>([1, 8, 188, 375])];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_183_cast_fp16 = reshape(shape = var_6473, x = var_6472_cast_fp16)[name = tensor<string, []>("matrix_bd_183_cast_fp16")];
+            tensor<int32, [4]> var_6478_begin_0 = const()[name = tensor<string, []>("op_6478_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_6478_end_0 = const()[name = tensor<string, []>("op_6478_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
+            tensor<bool, [4]> var_6478_end_mask_0 = const()[name = tensor<string, []>("op_6478_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 188, 188]> var_6478_cast_fp16 = slice_by_index(begin = var_6478_begin_0, end = var_6478_end_0, end_mask = var_6478_end_mask_0, x = matrix_bd_183_cast_fp16)[name = tensor<string, []>("op_6478_cast_fp16")];
+            tensor<fp16, []> var_6479_to_fp16 = const()[name = tensor<string, []>("op_6479_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 188, 188]> qk_mask_45_cast_fp16 = mul(x = var_6478_cast_fp16, y = var_6479_to_fp16)[name = tensor<string, []>("qk_mask_45_cast_fp16")];
+            tensor<int32, [4]> var_6483 = const()[name = tensor<string, []>("op_6483"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> mh_q_45_cast_fp16 = reshape(shape = var_6483, x = query_91_cast_fp16)[name = tensor<string, []>("mh_q_45_cast_fp16")];
+            tensor<fp16, []> var_6485_to_fp16 = const()[name = tensor<string, []>("op_6485_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 128, 188]> var_6486_cast_fp16 = mul(x = mh_q_45_cast_fp16, y = var_6485_to_fp16)[name = tensor<string, []>("op_6486_cast_fp16")];
+            tensor<int32, [4]> var_6489 = const()[name = tensor<string, []>("op_6489"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_6490_cast_fp16 = reshape(shape = var_6489, x = key_45_cast_fp16)[name = tensor<string, []>("op_6490_cast_fp16")];
             tensor<bool, []> mh_w_89_transpose_x_0 = const()[name = tensor<string, []>("mh_w_89_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> mh_w_89_transpose_y_0 = const()[name = tensor<string, []>("mh_w_89_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 188]> mh_w_89_cast_fp16 = matmul(transpose_x = mh_w_89_transpose_x_0, transpose_y = mh_w_89_transpose_y_0, x = var_6488_cast_fp16, y = var_6492_cast_fp16)[name = tensor<string, []>("mh_w_89_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> mh_w_89_cast_fp16 = matmul(transpose_x = mh_w_89_transpose_x_0, transpose_y = mh_w_89_transpose_y_0, x = var_6486_cast_fp16, y = var_6490_cast_fp16)[name = tensor<string, []>("mh_w_89_cast_fp16")];
             tensor<fp16, [1, 8, 188, 188]> mh_w_91_cast_fp16 = add(x = mh_w_89_cast_fp16, y = qk_mask_45_cast_fp16)[name = tensor<string, []>("mh_w_91_cast_fp16")];
-            tensor<fp16, [1, 8, 188, 188]> var_6496_cast_fp16 = softmax(axis = var_6337, x = mh_w_91_cast_fp16)[name = tensor<string, []>("op_6496_cast_fp16")];
-            tensor<int32, [4]> var_6497 = const()[name = tensor<string, []>("op_6497"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_6498_cast_fp16 = reshape(shape = var_6497, x = value_45_cast_fp16)[name = tensor<string, []>("op_6498_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> var_6494_cast_fp16 = softmax(axis = var_6335, x = mh_w_91_cast_fp16)[name = tensor<string, []>("op_6494_cast_fp16")];
+            tensor<int32, [4]> var_6495 = const()[name = tensor<string, []>("op_6495"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_6496_cast_fp16 = reshape(shape = var_6495, x = value_45_cast_fp16)[name = tensor<string, []>("op_6496_cast_fp16")];
             tensor<bool, []> attn_45_transpose_x_0 = const()[name = tensor<string, []>("attn_45_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> attn_45_transpose_y_0 = const()[name = tensor<string, []>("attn_45_transpose_y_0"), val = tensor<bool, []>(true)];
-            tensor<fp16, [1, 8, 128, 188]> attn_45_cast_fp16 = matmul(transpose_x = attn_45_transpose_x_0, transpose_y = attn_45_transpose_y_0, x = var_6498_cast_fp16, y = var_6496_cast_fp16)[name = tensor<string, []>("attn_45_cast_fp16")];
-            tensor<int32, [4]> var_6501 = const()[name = tensor<string, []>("op_6501"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
-            tensor<fp16, [1, 1024, 1, 188]> input_593_cast_fp16 = reshape(shape = var_6501, x = attn_45_cast_fp16)[name = tensor<string, []>("input_593_cast_fp16")];
+            tensor<fp16, [1, 8, 128, 188]> attn_45_cast_fp16 = matmul(transpose_x = attn_45_transpose_x_0, transpose_y = attn_45_transpose_y_0, x = var_6496_cast_fp16, y = var_6494_cast_fp16)[name = tensor<string, []>("attn_45_cast_fp16")];
+            tensor<int32, [4]> var_6499 = const()[name = tensor<string, []>("op_6499"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
+            tensor<fp16, [1, 1024, 1, 188]> input_595_cast_fp16 = reshape(shape = var_6499, x = attn_45_cast_fp16)[name = tensor<string, []>("input_595_cast_fp16")];
             tensor<string, []> obj_93_pad_type_0 = const()[name = tensor<string, []>("obj_93_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> obj_93_strides_0 = const()[name = tensor<string, []>("obj_93_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> obj_93_pad_0 = const()[name = tensor<string, []>("obj_93_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> obj_93_dilations_0 = const()[name = tensor<string, []>("obj_93_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> obj_93_groups_0 = const()[name = tensor<string, []>("obj_93_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_22_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_22_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1142952128)))];
-            tensor<fp16, [1, 1024, 1, 188]> obj_93_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = obj_93_dilations_0, groups = obj_93_groups_0, pad = obj_93_pad_0, pad_type = obj_93_pad_type_0, strides = obj_93_strides_0, weight = layers_22_self_attn_o_proj_weight_to_fp16, x = input_593_cast_fp16)[name = tensor<string, []>("obj_93_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_93_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_93_dilations_0, groups = obj_93_groups_0, pad = obj_93_pad_0, pad_type = obj_93_pad_type_0, strides = obj_93_strides_0, weight = layers_22_self_attn_o_proj_weight_to_fp16, x = input_595_cast_fp16)[name = tensor<string, []>("obj_93_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_225_cast_fp16 = add(x = inputs_223_cast_fp16, y = obj_93_cast_fp16)[name = tensor<string, []>("inputs_225_cast_fp16")];
             tensor<int32, [1]> out_225_axes_0 = const()[name = tensor<string, []>("out_225_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_6519_to_fp16 = const()[name = tensor<string, []>("op_6519_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_225_cast_fp16 = layer_norm(axes = out_225_axes_0, epsilon = var_6519_to_fp16, x = inputs_225_cast_fp16)[name = tensor<string, []>("out_225_cast_fp16")];
-            tensor<fp16, [1024]> input_595_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_595_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1145049344)))];
-            tensor<fp16, [1024]> input_595_beta_0_to_fp16 = const()[name = tensor<string, []>("input_595_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1145051456)))];
-            tensor<fp16, []> input_595_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_595_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_595_cast_fp16 = batch_norm(beta = input_595_beta_0_to_fp16, epsilon = input_595_epsilon_0_to_fp16, gamma = input_595_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_225_cast_fp16)[name = tensor<string, []>("input_595_cast_fp16")];
-            tensor<string, []> input_597_pad_type_0 = const()[name = tensor<string, []>("input_597_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_597_strides_0 = const()[name = tensor<string, []>("input_597_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_597_pad_0 = const()[name = tensor<string, []>("input_597_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_597_dilations_0 = const()[name = tensor<string, []>("input_597_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_597_groups_0 = const()[name = tensor<string, []>("input_597_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_6517_to_fp16 = const()[name = tensor<string, []>("op_6517_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_225_cast_fp16 = layer_norm(axes = out_225_axes_0, epsilon = var_6517_to_fp16, x = inputs_225_cast_fp16)[name = tensor<string, []>("out_225_cast_fp16")];
+            tensor<fp16, [1024]> input_597_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_597_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1145049344)))];
+            tensor<fp16, [1024]> input_597_beta_0_to_fp16 = const()[name = tensor<string, []>("input_597_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1145051456)))];
+            tensor<fp16, []> input_597_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_597_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_597_cast_fp16 = batch_norm(beta = input_597_beta_0_to_fp16, epsilon = input_597_epsilon_0_to_fp16, gamma = input_597_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_225_cast_fp16)[name = tensor<string, []>("input_597_cast_fp16")];
+            tensor<string, []> input_599_pad_type_0 = const()[name = tensor<string, []>("input_599_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_599_strides_0 = const()[name = tensor<string, []>("input_599_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_599_pad_0 = const()[name = tensor<string, []>("input_599_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_599_dilations_0 = const()[name = tensor<string, []>("input_599_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_599_groups_0 = const()[name = tensor<string, []>("input_599_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [2048, 1024, 1, 1]> layers_22_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("layers_22_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1145053568)))];
-            tensor<fp16, [1, 2048, 1, 188]> input_597_cast_fp16 = conv(dilations = input_597_dilations_0, groups = input_597_groups_0, pad = input_597_pad_0, pad_type = input_597_pad_type_0, strides = input_597_strides_0, weight = layers_22_conv_pointwise_conv1_weight_to_fp16, x = input_595_cast_fp16)[name = tensor<string, []>("input_597_cast_fp16")];
-            tensor<int32, []> input_599_split_num_splits_0 = const()[name = tensor<string, []>("input_599_split_num_splits_0"), val = tensor<int32, []>(2)];
-            tensor<int32, []> input_599_split_axis_0 = const()[name = tensor<string, []>("input_599_split_axis_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1, 1024, 1, 188]> input_599_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_599_split_cast_fp16_1 = split(axis = input_599_split_axis_0, num_splits = input_599_split_num_splits_0, x = input_597_cast_fp16)[name = tensor<string, []>("input_599_split_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_599_split_1_sigmoid_cast_fp16 = sigmoid(x = input_599_split_cast_fp16_1)[name = tensor<string, []>("input_599_split_1_sigmoid_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_599_cast_fp16 = mul(x = input_599_split_cast_fp16_0, y = input_599_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_599_cast_fp16")];
-            tensor<string, []> input_601_pad_type_0 = const()[name = tensor<string, []>("input_601_pad_type_0"), val = tensor<string, []>("custom")];
-            tensor<int32, [4]> input_601_pad_0 = const()[name = tensor<string, []>("input_601_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
-            tensor<int32, []> input_601_groups_0 = const()[name = tensor<string, []>("input_601_groups_0"), val = tensor<int32, []>(1024)];
-            tensor<int32, [2]> input_601_strides_0 = const()[name = tensor<string, []>("input_601_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> input_601_dilations_0 = const()[name = tensor<string, []>("input_601_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [1, 2048, 1, 188]> input_599_cast_fp16 = conv(dilations = input_599_dilations_0, groups = input_599_groups_0, pad = input_599_pad_0, pad_type = input_599_pad_type_0, strides = input_599_strides_0, weight = layers_22_conv_pointwise_conv1_weight_to_fp16, x = input_597_cast_fp16)[name = tensor<string, []>("input_599_cast_fp16")];
+            tensor<int32, []> input_601_split_num_splits_0 = const()[name = tensor<string, []>("input_601_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> input_601_split_axis_0 = const()[name = tensor<string, []>("input_601_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 1, 188]> input_601_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_601_split_cast_fp16_1 = split(axis = input_601_split_axis_0, num_splits = input_601_split_num_splits_0, x = input_599_cast_fp16)[name = tensor<string, []>("input_601_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_601_split_1_sigmoid_cast_fp16 = sigmoid(x = input_601_split_cast_fp16_1)[name = tensor<string, []>("input_601_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_601_cast_fp16 = mul(x = input_601_split_cast_fp16_0, y = input_601_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_601_cast_fp16")];
+            tensor<string, []> input_603_pad_type_0 = const()[name = tensor<string, []>("input_603_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_603_pad_0 = const()[name = tensor<string, []>("input_603_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
+            tensor<int32, []> input_603_groups_0 = const()[name = tensor<string, []>("input_603_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [2]> input_603_strides_0 = const()[name = tensor<string, []>("input_603_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> input_603_dilations_0 = const()[name = tensor<string, []>("input_603_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<fp16, [1024, 1, 1, 9]> const_312_to_fp16 = const()[name = tensor<string, []>("const_312_to_fp16"), val = tensor<fp16, [1024, 1, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1149247936)))];
             tensor<fp16, [1024]> const_313_to_fp16 = const()[name = tensor<string, []>("const_313_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1149266432)))];
-            tensor<fp16, [1, 1024, 1, 188]> input_603_cast_fp16 = conv(bias = const_313_to_fp16, dilations = input_601_dilations_0, groups = input_601_groups_0, pad = input_601_pad_0, pad_type = input_601_pad_type_0, strides = input_601_strides_0, weight = const_312_to_fp16, x = input_599_cast_fp16)[name = tensor<string, []>("input_603_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_605_cast_fp16 = silu(x = input_603_cast_fp16)[name = tensor<string, []>("input_605_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_605_cast_fp16 = conv(bias = const_313_to_fp16, dilations = input_603_dilations_0, groups = input_603_groups_0, pad = input_603_pad_0, pad_type = input_603_pad_type_0, strides = input_603_strides_0, weight = const_312_to_fp16, x = input_601_cast_fp16)[name = tensor<string, []>("input_605_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_607_cast_fp16 = silu(x = input_605_cast_fp16)[name = tensor<string, []>("input_607_cast_fp16")];
             tensor<string, []> x_137_pad_type_0 = const()[name = tensor<string, []>("x_137_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_137_strides_0 = const()[name = tensor<string, []>("x_137_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_137_pad_0 = const()[name = tensor<string, []>("x_137_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_137_dilations_0 = const()[name = tensor<string, []>("x_137_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_137_groups_0 = const()[name = tensor<string, []>("x_137_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_22_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("layers_22_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1149268544)))];
-            tensor<fp16, [1, 1024, 1, 188]> x_137_cast_fp16 = conv(dilations = x_137_dilations_0, groups = x_137_groups_0, pad = x_137_pad_0, pad_type = x_137_pad_type_0, strides = x_137_strides_0, weight = layers_22_conv_pointwise_conv2_weight_to_fp16, x = input_605_cast_fp16)[name = tensor<string, []>("x_137_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> x_137_cast_fp16 = conv(dilations = x_137_dilations_0, groups = x_137_groups_0, pad = x_137_pad_0, pad_type = x_137_pad_type_0, strides = x_137_strides_0, weight = layers_22_conv_pointwise_conv2_weight_to_fp16, x = input_607_cast_fp16)[name = tensor<string, []>("x_137_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_227_cast_fp16 = add(x = inputs_225_cast_fp16, y = x_137_cast_fp16)[name = tensor<string, []>("inputs_227_cast_fp16")];
             tensor<int32, [1]> out_227_axes_0 = const()[name = tensor<string, []>("out_227_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_6567_to_fp16 = const()[name = tensor<string, []>("op_6567_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_227_cast_fp16 = layer_norm(axes = out_227_axes_0, epsilon = var_6567_to_fp16, x = inputs_227_cast_fp16)[name = tensor<string, []>("out_227_cast_fp16")];
-            tensor<fp16, [1024]> input_607_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_607_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1151365760)))];
-            tensor<fp16, [1024]> input_607_beta_0_to_fp16 = const()[name = tensor<string, []>("input_607_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1151367872)))];
-            tensor<fp16, []> input_607_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_607_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_607_cast_fp16 = batch_norm(beta = input_607_beta_0_to_fp16, epsilon = input_607_epsilon_0_to_fp16, gamma = input_607_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_227_cast_fp16)[name = tensor<string, []>("input_607_cast_fp16")];
-            tensor<string, []> input_609_pad_type_0 = const()[name = tensor<string, []>("input_609_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_609_strides_0 = const()[name = tensor<string, []>("input_609_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_609_pad_0 = const()[name = tensor<string, []>("input_609_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_609_dilations_0 = const()[name = tensor<string, []>("input_609_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_609_groups_0 = const()[name = tensor<string, []>("input_609_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_6565_to_fp16 = const()[name = tensor<string, []>("op_6565_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_227_cast_fp16 = layer_norm(axes = out_227_axes_0, epsilon = var_6565_to_fp16, x = inputs_227_cast_fp16)[name = tensor<string, []>("out_227_cast_fp16")];
+            tensor<fp16, [1024]> input_609_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_609_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1151365760)))];
+            tensor<fp16, [1024]> input_609_beta_0_to_fp16 = const()[name = tensor<string, []>("input_609_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1151367872)))];
+            tensor<fp16, []> input_609_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_609_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_609_cast_fp16 = batch_norm(beta = input_609_beta_0_to_fp16, epsilon = input_609_epsilon_0_to_fp16, gamma = input_609_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_227_cast_fp16)[name = tensor<string, []>("input_609_cast_fp16")];
+            tensor<string, []> input_611_pad_type_0 = const()[name = tensor<string, []>("input_611_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_611_strides_0 = const()[name = tensor<string, []>("input_611_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_611_pad_0 = const()[name = tensor<string, []>("input_611_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_611_dilations_0 = const()[name = tensor<string, []>("input_611_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_611_groups_0 = const()[name = tensor<string, []>("input_611_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_22_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_22_feed_forward2_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1151369984)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_609_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_609_dilations_0, groups = input_609_groups_0, pad = input_609_pad_0, pad_type = input_609_pad_type_0, strides = input_609_strides_0, weight = layers_22_feed_forward2_fc1_weight_to_fp16, x = input_607_cast_fp16)[name = tensor<string, []>("input_609_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_611_cast_fp16 = silu(x = input_609_cast_fp16)[name = tensor<string, []>("input_611_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_611_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_611_dilations_0, groups = input_611_groups_0, pad = input_611_pad_0, pad_type = input_611_pad_type_0, strides = input_611_strides_0, weight = layers_22_feed_forward2_fc1_weight_to_fp16, x = input_609_cast_fp16)[name = tensor<string, []>("input_611_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_613_cast_fp16 = silu(x = input_611_cast_fp16)[name = tensor<string, []>("input_613_cast_fp16")];
             tensor<string, []> x_139_pad_type_0 = const()[name = tensor<string, []>("x_139_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_139_strides_0 = const()[name = tensor<string, []>("x_139_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_139_pad_0 = const()[name = tensor<string, []>("x_139_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_139_dilations_0 = const()[name = tensor<string, []>("x_139_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_139_groups_0 = const()[name = tensor<string, []>("x_139_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_6595_weight_0_to_fp16 = const()[name = tensor<string, []>("op_6595_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1159758656)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_6595_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_139_dilations_0, groups = x_139_groups_0, pad = x_139_pad_0, pad_type = x_139_pad_type_0, strides = x_139_strides_0, weight = var_6595_weight_0_to_fp16, x = input_611_cast_fp16)[name = tensor<string, []>("op_6595_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_229_cast_fp16 = add(x = inputs_227_cast_fp16, y = var_6595_cast_fp16)[name = tensor<string, []>("inputs_229_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_6593_weight_0_to_fp16 = const()[name = tensor<string, []>("op_6593_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1159758656)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_6593_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_139_dilations_0, groups = x_139_groups_0, pad = x_139_pad_0, pad_type = x_139_pad_type_0, strides = x_139_strides_0, weight = var_6593_weight_0_to_fp16, x = input_613_cast_fp16)[name = tensor<string, []>("op_6593_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_229_cast_fp16 = add(x = inputs_227_cast_fp16, y = var_6593_cast_fp16)[name = tensor<string, []>("inputs_229_cast_fp16")];
             tensor<int32, [1]> out_229_axes_0 = const()[name = tensor<string, []>("out_229_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_6605_to_fp16 = const()[name = tensor<string, []>("op_6605_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_229_cast_fp16 = layer_norm(axes = out_229_axes_0, epsilon = var_6605_to_fp16, x = inputs_229_cast_fp16)[name = tensor<string, []>("out_229_cast_fp16")];
+            tensor<fp16, []> var_6603_to_fp16 = const()[name = tensor<string, []>("op_6603_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_229_cast_fp16 = layer_norm(axes = out_229_axes_0, epsilon = var_6603_to_fp16, x = inputs_229_cast_fp16)[name = tensor<string, []>("out_229_cast_fp16")];
             tensor<fp16, [1024]> inputs_231_gamma_0_to_fp16 = const()[name = tensor<string, []>("inputs_231_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1168147328)))];
             tensor<fp16, [1024]> inputs_231_beta_0_to_fp16 = const()[name = tensor<string, []>("inputs_231_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1168149440)))];
             tensor<fp16, []> inputs_231_epsilon_0_to_fp16 = const()[name = tensor<string, []>("inputs_231_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_231_cast_fp16 = batch_norm(beta = inputs_231_beta_0_to_fp16, epsilon = inputs_231_epsilon_0_to_fp16, gamma = inputs_231_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_229_cast_fp16)[name = tensor<string, []>("inputs_231_cast_fp16")];
-            tensor<int32, []> var_6619 = const()[name = tensor<string, []>("op_6619"), val = tensor<int32, []>(3)];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_231_cast_fp16 = batch_norm(beta = inputs_231_beta_0_to_fp16, epsilon = inputs_231_epsilon_0_to_fp16, gamma = inputs_231_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_229_cast_fp16)[name = tensor<string, []>("inputs_231_cast_fp16")];
+            tensor<int32, []> var_6617 = const()[name = tensor<string, []>("op_6617"), val = tensor<int32, []>(3)];
             tensor<int32, [1]> out_231_axes_0 = const()[name = tensor<string, []>("out_231_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_6650_to_fp16 = const()[name = tensor<string, []>("op_6650_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_231_cast_fp16 = layer_norm(axes = out_231_axes_0, epsilon = var_6650_to_fp16, x = inputs_231_cast_fp16)[name = tensor<string, []>("out_231_cast_fp16")];
-            tensor<fp16, [1024]> input_613_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_613_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1168151552)))];
-            tensor<fp16, [1024]> input_613_beta_0_to_fp16 = const()[name = tensor<string, []>("input_613_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1168153664)))];
-            tensor<fp16, []> input_613_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_613_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_613_cast_fp16 = batch_norm(beta = input_613_beta_0_to_fp16, epsilon = input_613_epsilon_0_to_fp16, gamma = input_613_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_231_cast_fp16)[name = tensor<string, []>("input_613_cast_fp16")];
-            tensor<string, []> input_615_pad_type_0 = const()[name = tensor<string, []>("input_615_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_615_strides_0 = const()[name = tensor<string, []>("input_615_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_615_pad_0 = const()[name = tensor<string, []>("input_615_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_615_dilations_0 = const()[name = tensor<string, []>("input_615_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_615_groups_0 = const()[name = tensor<string, []>("input_615_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_6648_to_fp16 = const()[name = tensor<string, []>("op_6648_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_231_cast_fp16 = layer_norm(axes = out_231_axes_0, epsilon = var_6648_to_fp16, x = inputs_231_cast_fp16)[name = tensor<string, []>("out_231_cast_fp16")];
+            tensor<fp16, [1024]> input_615_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_615_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1168151552)))];
+            tensor<fp16, [1024]> input_615_beta_0_to_fp16 = const()[name = tensor<string, []>("input_615_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1168153664)))];
+            tensor<fp16, []> input_615_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_615_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_615_cast_fp16 = batch_norm(beta = input_615_beta_0_to_fp16, epsilon = input_615_epsilon_0_to_fp16, gamma = input_615_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_231_cast_fp16)[name = tensor<string, []>("input_615_cast_fp16")];
+            tensor<string, []> input_617_pad_type_0 = const()[name = tensor<string, []>("input_617_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_617_strides_0 = const()[name = tensor<string, []>("input_617_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_617_pad_0 = const()[name = tensor<string, []>("input_617_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_617_dilations_0 = const()[name = tensor<string, []>("input_617_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_617_groups_0 = const()[name = tensor<string, []>("input_617_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_23_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_23_feed_forward1_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1168155776)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_615_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_615_dilations_0, groups = input_615_groups_0, pad = input_615_pad_0, pad_type = input_615_pad_type_0, strides = input_615_strides_0, weight = layers_23_feed_forward1_fc1_weight_to_fp16, x = input_613_cast_fp16)[name = tensor<string, []>("input_615_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_617_cast_fp16 = silu(x = input_615_cast_fp16)[name = tensor<string, []>("input_617_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_617_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_617_dilations_0, groups = input_617_groups_0, pad = input_617_pad_0, pad_type = input_617_pad_type_0, strides = input_617_strides_0, weight = layers_23_feed_forward1_fc1_weight_to_fp16, x = input_615_cast_fp16)[name = tensor<string, []>("input_617_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_619_cast_fp16 = silu(x = input_617_cast_fp16)[name = tensor<string, []>("input_619_cast_fp16")];
             tensor<string, []> x_141_pad_type_0 = const()[name = tensor<string, []>("x_141_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_141_strides_0 = const()[name = tensor<string, []>("x_141_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_141_pad_0 = const()[name = tensor<string, []>("x_141_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_141_dilations_0 = const()[name = tensor<string, []>("x_141_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_141_groups_0 = const()[name = tensor<string, []>("x_141_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_6678_weight_0_to_fp16 = const()[name = tensor<string, []>("op_6678_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1176544448)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_6678_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_141_dilations_0, groups = x_141_groups_0, pad = x_141_pad_0, pad_type = x_141_pad_type_0, strides = x_141_strides_0, weight = var_6678_weight_0_to_fp16, x = input_617_cast_fp16)[name = tensor<string, []>("op_6678_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_233_cast_fp16 = add(x = inputs_231_cast_fp16, y = var_6678_cast_fp16)[name = tensor<string, []>("inputs_233_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_6676_weight_0_to_fp16 = const()[name = tensor<string, []>("op_6676_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1176544448)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_6676_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_141_dilations_0, groups = x_141_groups_0, pad = x_141_pad_0, pad_type = x_141_pad_type_0, strides = x_141_strides_0, weight = var_6676_weight_0_to_fp16, x = input_619_cast_fp16)[name = tensor<string, []>("op_6676_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_233_cast_fp16 = add(x = inputs_231_cast_fp16, y = var_6676_cast_fp16)[name = tensor<string, []>("inputs_233_cast_fp16")];
             tensor<int32, [1]> out_233_axes_0 = const()[name = tensor<string, []>("out_233_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_6688_to_fp16 = const()[name = tensor<string, []>("op_6688_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_233_cast_fp16 = layer_norm(axes = out_233_axes_0, epsilon = var_6688_to_fp16, x = inputs_233_cast_fp16)[name = tensor<string, []>("out_233_cast_fp16")];
+            tensor<fp16, []> var_6686_to_fp16 = const()[name = tensor<string, []>("op_6686_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_233_cast_fp16 = layer_norm(axes = out_233_axes_0, epsilon = var_6686_to_fp16, x = inputs_233_cast_fp16)[name = tensor<string, []>("out_233_cast_fp16")];
             tensor<fp16, [1024]> obj_95_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_95_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1184933120)))];
             tensor<fp16, [1024]> obj_95_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_95_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1184935232)))];
             tensor<fp16, []> obj_95_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_95_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> obj_95_cast_fp16 = batch_norm(beta = obj_95_beta_0_to_fp16, epsilon = obj_95_epsilon_0_to_fp16, gamma = obj_95_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_233_cast_fp16)[name = tensor<string, []>("obj_95_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_95_cast_fp16 = batch_norm(beta = obj_95_beta_0_to_fp16, epsilon = obj_95_epsilon_0_to_fp16, gamma = obj_95_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_233_cast_fp16)[name = tensor<string, []>("obj_95_cast_fp16")];
             tensor<string, []> query_93_pad_type_0 = const()[name = tensor<string, []>("query_93_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> query_93_strides_0 = const()[name = tensor<string, []>("query_93_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> query_93_pad_0 = const()[name = tensor<string, []>("query_93_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> query_93_dilations_0 = const()[name = tensor<string, []>("query_93_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> query_93_groups_0 = const()[name = tensor<string, []>("query_93_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_23_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_23_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1184937344)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_93_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = query_93_dilations_0, groups = query_93_groups_0, pad = query_93_pad_0, pad_type = query_93_pad_type_0, strides = query_93_strides_0, weight = layers_23_self_attn_q_proj_weight_to_fp16, x = obj_95_cast_fp16)[name = tensor<string, []>("query_93_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> query_93_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_93_dilations_0, groups = query_93_groups_0, pad = query_93_pad_0, pad_type = query_93_pad_type_0, strides = query_93_strides_0, weight = layers_23_self_attn_q_proj_weight_to_fp16, x = obj_95_cast_fp16)[name = tensor<string, []>("query_93_cast_fp16")];
             tensor<string, []> key_pad_type_0 = const()[name = tensor<string, []>("key_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> key_strides_0 = const()[name = tensor<string, []>("key_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> key_pad_0 = const()[name = tensor<string, []>("key_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
@@ -4252,142 +4254,142 @@ program(1.0)
             tensor<int32, [2]> value_dilations_0 = const()[name = tensor<string, []>("value_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> value_groups_0 = const()[name = tensor<string, []>("value_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_23_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_23_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1189131776)))];
-            tensor<fp16, [1, 1024, 1, 188]> value_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = value_dilations_0, groups = value_groups_0, pad = value_pad_0, pad_type = value_pad_type_0, strides = value_strides_0, weight = layers_23_self_attn_v_proj_weight_to_fp16, x = obj_95_cast_fp16)[name = tensor<string, []>("value_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_6726_to_fp16 = const()[name = tensor<string, []>("op_6726_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1191228992)))];
-            tensor<fp16, [1, 1024, 1, 188]> query_cast_fp16 = add(x = query_93_cast_fp16, y = var_6726_to_fp16)[name = tensor<string, []>("query_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 1]> var_6729_to_fp16 = const()[name = tensor<string, []>("op_6729_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1191231104)))];
-            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_cast_fp16 = add(x = query_93_cast_fp16, y = var_6729_to_fp16)[name = tensor<string, []>("q_with_bias_v_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> value_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_dilations_0, groups = value_groups_0, pad = value_pad_0, pad_type = value_pad_type_0, strides = value_strides_0, weight = layers_23_self_attn_v_proj_weight_to_fp16, x = obj_95_cast_fp16)[name = tensor<string, []>("value_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_6724_to_fp16 = const()[name = tensor<string, []>("op_6724_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1191228992)))];
+            tensor<fp16, [1, 1024, 1, 188]> query_cast_fp16 = add(x = query_93_cast_fp16, y = var_6724_to_fp16)[name = tensor<string, []>("query_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 1]> var_6727_to_fp16 = const()[name = tensor<string, []>("op_6727_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1191231104)))];
+            tensor<fp16, [1, 1024, 1, 188]> q_with_bias_v_cast_fp16 = add(x = query_93_cast_fp16, y = var_6727_to_fp16)[name = tensor<string, []>("q_with_bias_v_cast_fp16")];
             tensor<string, []> p_pad_type_0 = const()[name = tensor<string, []>("p_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> p_strides_0 = const()[name = tensor<string, []>("p_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> p_pad_0 = const()[name = tensor<string, []>("p_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> p_dilations_0 = const()[name = tensor<string, []>("p_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> p_groups_0 = const()[name = tensor<string, []>("p_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_23_self_attn_linear_pos_weight_to_fp16 = const()[name = tensor<string, []>("layers_23_self_attn_linear_pos_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1191233216)))];
-            tensor<fp16, [1, 1024, 1, 375]> p_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = p_dilations_0, groups = p_groups_0, pad = p_pad_0, pad_type = p_pad_type_0, strides = p_strides_0, weight = layers_23_self_attn_linear_pos_weight_to_fp16, x = pos_enc_to_fp16)[name = tensor<string, []>("p_cast_fp16")];
-            tensor<int32, [4]> var_6740 = const()[name = tensor<string, []>("op_6740"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_6741_cast_fp16 = reshape(shape = var_6740, x = q_with_bias_v_cast_fp16)[name = tensor<string, []>("op_6741_cast_fp16")];
-            tensor<int32, [4]> var_6742 = const()[name = tensor<string, []>("op_6742"), val = tensor<int32, [4]>([1, 8, 128, -1])];
-            tensor<fp16, [1, 8, 128, 375]> var_6743_cast_fp16 = reshape(shape = var_6742, x = p_cast_fp16)[name = tensor<string, []>("op_6743_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 375]> p_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_dilations_0, groups = p_groups_0, pad = p_pad_0, pad_type = p_pad_type_0, strides = p_strides_0, weight = layers_23_self_attn_linear_pos_weight_to_fp16, x = obj_3_cast_fp16)[name = tensor<string, []>("p_cast_fp16")];
+            tensor<int32, [4]> var_6738 = const()[name = tensor<string, []>("op_6738"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_6739_cast_fp16 = reshape(shape = var_6738, x = q_with_bias_v_cast_fp16)[name = tensor<string, []>("op_6739_cast_fp16")];
+            tensor<int32, [4]> var_6740 = const()[name = tensor<string, []>("op_6740"), val = tensor<int32, [4]>([1, 8, 128, -1])];
+            tensor<fp16, [1, 8, 128, 375]> var_6741_cast_fp16 = reshape(shape = var_6740, x = p_cast_fp16)[name = tensor<string, []>("op_6741_cast_fp16")];
             tensor<bool, []> matrix_bd_185_transpose_x_0 = const()[name = tensor<string, []>("matrix_bd_185_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> matrix_bd_185_transpose_y_0 = const()[name = tensor<string, []>("matrix_bd_185_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_185_cast_fp16 = matmul(transpose_x = matrix_bd_185_transpose_x_0, transpose_y = matrix_bd_185_transpose_y_0, x = var_6741_cast_fp16, y = var_6743_cast_fp16)[name = tensor<string, []>("matrix_bd_185_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_185_cast_fp16 = matmul(transpose_x = matrix_bd_185_transpose_x_0, transpose_y = matrix_bd_185_transpose_y_0, x = var_6739_cast_fp16, y = var_6741_cast_fp16)[name = tensor<string, []>("matrix_bd_185_cast_fp16")];
             tensor<int32, [8]> matrix_bd_187_pad_0 = const()[name = tensor<string, []>("matrix_bd_187_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
             tensor<string, []> matrix_bd_187_mode_0 = const()[name = tensor<string, []>("matrix_bd_187_mode_0"), val = tensor<string, []>("constant")];
             tensor<fp16, []> const_263_to_fp16 = const()[name = tensor<string, []>("const_263_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
             tensor<fp16, [1, 8, 188, 376]> matrix_bd_187_cast_fp16 = pad(constant_val = const_263_to_fp16, mode = matrix_bd_187_mode_0, pad = matrix_bd_187_pad_0, x = matrix_bd_185_cast_fp16)[name = tensor<string, []>("matrix_bd_187_cast_fp16")];
-            tensor<int32, [4]> var_6752 = const()[name = tensor<string, []>("op_6752"), val = tensor<int32, [4]>([1, 8, -1, 188])];
-            tensor<fp16, [1, 8, 376, 188]> matrix_bd_189_cast_fp16 = reshape(shape = var_6752, x = matrix_bd_187_cast_fp16)[name = tensor<string, []>("matrix_bd_189_cast_fp16")];
-            tensor<int32, [4]> var_6756_begin_0 = const()[name = tensor<string, []>("op_6756_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
-            tensor<int32, [4]> var_6756_end_0 = const()[name = tensor<string, []>("op_6756_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
-            tensor<bool, [4]> var_6756_end_mask_0 = const()[name = tensor<string, []>("op_6756_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
-            tensor<fp16, [1, 8, 375, 188]> var_6756_cast_fp16 = slice_by_index(begin = var_6756_begin_0, end = var_6756_end_0, end_mask = var_6756_end_mask_0, x = matrix_bd_189_cast_fp16)[name = tensor<string, []>("op_6756_cast_fp16")];
-            tensor<int32, [4]> var_6757 = const()[name = tensor<string, []>("op_6757"), val = tensor<int32, [4]>([1, 8, 188, 375])];
-            tensor<fp16, [1, 8, 188, 375]> matrix_bd_cast_fp16 = reshape(shape = var_6757, x = var_6756_cast_fp16)[name = tensor<string, []>("matrix_bd_cast_fp16")];
-            tensor<int32, [4]> var_6762_begin_0 = const()[name = tensor<string, []>("op_6762_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [4]> var_6762_end_0 = const()[name = tensor<string, []>("op_6762_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
-            tensor<bool, [4]> var_6762_end_mask_0 = const()[name = tensor<string, []>("op_6762_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
-            tensor<fp16, [1, 8, 188, 188]> var_6762_cast_fp16 = slice_by_index(begin = var_6762_begin_0, end = var_6762_end_0, end_mask = var_6762_end_mask_0, x = matrix_bd_cast_fp16)[name = tensor<string, []>("op_6762_cast_fp16")];
-            tensor<fp16, []> var_6763_to_fp16 = const()[name = tensor<string, []>("op_6763_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 188, 188]> qk_mask_cast_fp16 = mul(x = var_6762_cast_fp16, y = var_6763_to_fp16)[name = tensor<string, []>("qk_mask_cast_fp16")];
-            tensor<int32, [4]> var_6767 = const()[name = tensor<string, []>("op_6767"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> mh_q_cast_fp16 = reshape(shape = var_6767, x = query_cast_fp16)[name = tensor<string, []>("mh_q_cast_fp16")];
-            tensor<fp16, []> var_6769_to_fp16 = const()[name = tensor<string, []>("op_6769_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
-            tensor<fp16, [1, 8, 128, 188]> var_6770_cast_fp16 = mul(x = mh_q_cast_fp16, y = var_6769_to_fp16)[name = tensor<string, []>("op_6770_cast_fp16")];
-            tensor<int32, [4]> var_6773 = const()[name = tensor<string, []>("op_6773"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_6774_cast_fp16 = reshape(shape = var_6773, x = key_cast_fp16)[name = tensor<string, []>("op_6774_cast_fp16")];
+            tensor<int32, [4]> var_6750 = const()[name = tensor<string, []>("op_6750"), val = tensor<int32, [4]>([1, 8, -1, 188])];
+            tensor<fp16, [1, 8, 376, 188]> matrix_bd_189_cast_fp16 = reshape(shape = var_6750, x = matrix_bd_187_cast_fp16)[name = tensor<string, []>("matrix_bd_189_cast_fp16")];
+            tensor<int32, [4]> var_6754_begin_0 = const()[name = tensor<string, []>("op_6754_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_6754_end_0 = const()[name = tensor<string, []>("op_6754_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
+            tensor<bool, [4]> var_6754_end_mask_0 = const()[name = tensor<string, []>("op_6754_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 375, 188]> var_6754_cast_fp16 = slice_by_index(begin = var_6754_begin_0, end = var_6754_end_0, end_mask = var_6754_end_mask_0, x = matrix_bd_189_cast_fp16)[name = tensor<string, []>("op_6754_cast_fp16")];
+            tensor<int32, [4]> var_6755 = const()[name = tensor<string, []>("op_6755"), val = tensor<int32, [4]>([1, 8, 188, 375])];
+            tensor<fp16, [1, 8, 188, 375]> matrix_bd_cast_fp16 = reshape(shape = var_6755, x = var_6754_cast_fp16)[name = tensor<string, []>("matrix_bd_cast_fp16")];
+            tensor<int32, [4]> var_6760_begin_0 = const()[name = tensor<string, []>("op_6760_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_6760_end_0 = const()[name = tensor<string, []>("op_6760_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
+            tensor<bool, [4]> var_6760_end_mask_0 = const()[name = tensor<string, []>("op_6760_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 188, 188]> var_6760_cast_fp16 = slice_by_index(begin = var_6760_begin_0, end = var_6760_end_0, end_mask = var_6760_end_mask_0, x = matrix_bd_cast_fp16)[name = tensor<string, []>("op_6760_cast_fp16")];
+            tensor<fp16, []> var_6761_to_fp16 = const()[name = tensor<string, []>("op_6761_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 188, 188]> qk_mask_cast_fp16 = mul(x = var_6760_cast_fp16, y = var_6761_to_fp16)[name = tensor<string, []>("qk_mask_cast_fp16")];
+            tensor<int32, [4]> var_6765 = const()[name = tensor<string, []>("op_6765"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> mh_q_cast_fp16 = reshape(shape = var_6765, x = query_cast_fp16)[name = tensor<string, []>("mh_q_cast_fp16")];
+            tensor<fp16, []> var_6767_to_fp16 = const()[name = tensor<string, []>("op_6767_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 128, 188]> var_6768_cast_fp16 = mul(x = mh_q_cast_fp16, y = var_6767_to_fp16)[name = tensor<string, []>("op_6768_cast_fp16")];
+            tensor<int32, [4]> var_6771 = const()[name = tensor<string, []>("op_6771"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_6772_cast_fp16 = reshape(shape = var_6771, x = key_cast_fp16)[name = tensor<string, []>("op_6772_cast_fp16")];
             tensor<bool, []> mh_w_93_transpose_x_0 = const()[name = tensor<string, []>("mh_w_93_transpose_x_0"), val = tensor<bool, []>(true)];
             tensor<bool, []> mh_w_93_transpose_y_0 = const()[name = tensor<string, []>("mh_w_93_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 188, 188]> mh_w_93_cast_fp16 = matmul(transpose_x = mh_w_93_transpose_x_0, transpose_y = mh_w_93_transpose_y_0, x = var_6770_cast_fp16, y = var_6774_cast_fp16)[name = tensor<string, []>("mh_w_93_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> mh_w_93_cast_fp16 = matmul(transpose_x = mh_w_93_transpose_x_0, transpose_y = mh_w_93_transpose_y_0, x = var_6768_cast_fp16, y = var_6772_cast_fp16)[name = tensor<string, []>("mh_w_93_cast_fp16")];
             tensor<fp16, [1, 8, 188, 188]> mh_w_cast_fp16 = add(x = mh_w_93_cast_fp16, y = qk_mask_cast_fp16)[name = tensor<string, []>("mh_w_cast_fp16")];
-            tensor<fp16, [1, 8, 188, 188]> var_6778_cast_fp16 = softmax(axis = var_6619, x = mh_w_cast_fp16)[name = tensor<string, []>("op_6778_cast_fp16")];
-            tensor<int32, [4]> var_6779 = const()[name = tensor<string, []>("op_6779"), val = tensor<int32, [4]>([1, 8, 128, 188])];
-            tensor<fp16, [1, 8, 128, 188]> var_6780_cast_fp16 = reshape(shape = var_6779, x = value_cast_fp16)[name = tensor<string, []>("op_6780_cast_fp16")];
+            tensor<fp16, [1, 8, 188, 188]> var_6776_cast_fp16 = softmax(axis = var_6617, x = mh_w_cast_fp16)[name = tensor<string, []>("op_6776_cast_fp16")];
+            tensor<int32, [4]> var_6777 = const()[name = tensor<string, []>("op_6777"), val = tensor<int32, [4]>([1, 8, 128, 188])];
+            tensor<fp16, [1, 8, 128, 188]> var_6778_cast_fp16 = reshape(shape = var_6777, x = value_cast_fp16)[name = tensor<string, []>("op_6778_cast_fp16")];
             tensor<bool, []> attn_transpose_x_0 = const()[name = tensor<string, []>("attn_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> attn_transpose_y_0 = const()[name = tensor<string, []>("attn_transpose_y_0"), val = tensor<bool, []>(true)];
-            tensor<fp16, [1, 8, 128, 188]> attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = var_6780_cast_fp16, y = var_6778_cast_fp16)[name = tensor<string, []>("attn_cast_fp16")];
-            tensor<int32, [4]> var_6783 = const()[name = tensor<string, []>("op_6783"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
-            tensor<fp16, [1, 1024, 1, 188]> input_619_cast_fp16 = reshape(shape = var_6783, x = attn_cast_fp16)[name = tensor<string, []>("input_619_cast_fp16")];
+            tensor<fp16, [1, 8, 128, 188]> attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = var_6778_cast_fp16, y = var_6776_cast_fp16)[name = tensor<string, []>("attn_cast_fp16")];
+            tensor<int32, [4]> var_6781 = const()[name = tensor<string, []>("op_6781"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
+            tensor<fp16, [1, 1024, 1, 188]> input_621_cast_fp16 = reshape(shape = var_6781, x = attn_cast_fp16)[name = tensor<string, []>("input_621_cast_fp16")];
             tensor<string, []> obj_pad_type_0 = const()[name = tensor<string, []>("obj_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> obj_strides_0 = const()[name = tensor<string, []>("obj_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> obj_pad_0 = const()[name = tensor<string, []>("obj_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> obj_dilations_0 = const()[name = tensor<string, []>("obj_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> obj_groups_0 = const()[name = tensor<string, []>("obj_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_23_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_23_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1193330432)))];
-            tensor<fp16, [1, 1024, 1, 188]> obj_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = obj_dilations_0, groups = obj_groups_0, pad = obj_pad_0, pad_type = obj_pad_type_0, strides = obj_strides_0, weight = layers_23_self_attn_o_proj_weight_to_fp16, x = input_619_cast_fp16)[name = tensor<string, []>("obj_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> obj_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_dilations_0, groups = obj_groups_0, pad = obj_pad_0, pad_type = obj_pad_type_0, strides = obj_strides_0, weight = layers_23_self_attn_o_proj_weight_to_fp16, x = input_621_cast_fp16)[name = tensor<string, []>("obj_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_235_cast_fp16 = add(x = inputs_233_cast_fp16, y = obj_cast_fp16)[name = tensor<string, []>("inputs_235_cast_fp16")];
             tensor<int32, [1]> out_235_axes_0 = const()[name = tensor<string, []>("out_235_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_6801_to_fp16 = const()[name = tensor<string, []>("op_6801_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_235_cast_fp16 = layer_norm(axes = out_235_axes_0, epsilon = var_6801_to_fp16, x = inputs_235_cast_fp16)[name = tensor<string, []>("out_235_cast_fp16")];
-            tensor<fp16, [1024]> input_621_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_621_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1195427648)))];
-            tensor<fp16, [1024]> input_621_beta_0_to_fp16 = const()[name = tensor<string, []>("input_621_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1195429760)))];
-            tensor<fp16, []> input_621_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_621_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_621_cast_fp16 = batch_norm(beta = input_621_beta_0_to_fp16, epsilon = input_621_epsilon_0_to_fp16, gamma = input_621_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_235_cast_fp16)[name = tensor<string, []>("input_621_cast_fp16")];
-            tensor<string, []> input_623_pad_type_0 = const()[name = tensor<string, []>("input_623_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_623_strides_0 = const()[name = tensor<string, []>("input_623_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_623_pad_0 = const()[name = tensor<string, []>("input_623_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_623_dilations_0 = const()[name = tensor<string, []>("input_623_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_623_groups_0 = const()[name = tensor<string, []>("input_623_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_6799_to_fp16 = const()[name = tensor<string, []>("op_6799_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_235_cast_fp16 = layer_norm(axes = out_235_axes_0, epsilon = var_6799_to_fp16, x = inputs_235_cast_fp16)[name = tensor<string, []>("out_235_cast_fp16")];
+            tensor<fp16, [1024]> input_623_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_623_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1195427648)))];
+            tensor<fp16, [1024]> input_623_beta_0_to_fp16 = const()[name = tensor<string, []>("input_623_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1195429760)))];
+            tensor<fp16, []> input_623_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_623_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_623_cast_fp16 = batch_norm(beta = input_623_beta_0_to_fp16, epsilon = input_623_epsilon_0_to_fp16, gamma = input_623_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_235_cast_fp16)[name = tensor<string, []>("input_623_cast_fp16")];
+            tensor<string, []> input_625_pad_type_0 = const()[name = tensor<string, []>("input_625_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_625_strides_0 = const()[name = tensor<string, []>("input_625_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_625_pad_0 = const()[name = tensor<string, []>("input_625_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_625_dilations_0 = const()[name = tensor<string, []>("input_625_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_625_groups_0 = const()[name = tensor<string, []>("input_625_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [2048, 1024, 1, 1]> layers_23_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor<string, []>("layers_23_conv_pointwise_conv1_weight_to_fp16"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1195431872)))];
-            tensor<fp16, [1, 2048, 1, 188]> input_623_cast_fp16 = conv(dilations = input_623_dilations_0, groups = input_623_groups_0, pad = input_623_pad_0, pad_type = input_623_pad_type_0, strides = input_623_strides_0, weight = layers_23_conv_pointwise_conv1_weight_to_fp16, x = input_621_cast_fp16)[name = tensor<string, []>("input_623_cast_fp16")];
-            tensor<int32, []> input_625_split_num_splits_0 = const()[name = tensor<string, []>("input_625_split_num_splits_0"), val = tensor<int32, []>(2)];
-            tensor<int32, []> input_625_split_axis_0 = const()[name = tensor<string, []>("input_625_split_axis_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1, 1024, 1, 188]> input_625_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_625_split_cast_fp16_1 = split(axis = input_625_split_axis_0, num_splits = input_625_split_num_splits_0, x = input_623_cast_fp16)[name = tensor<string, []>("input_625_split_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_625_split_1_sigmoid_cast_fp16 = sigmoid(x = input_625_split_cast_fp16_1)[name = tensor<string, []>("input_625_split_1_sigmoid_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_625_cast_fp16 = mul(x = input_625_split_cast_fp16_0, y = input_625_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_625_cast_fp16")];
-            tensor<string, []> input_627_pad_type_0 = const()[name = tensor<string, []>("input_627_pad_type_0"), val = tensor<string, []>("custom")];
-            tensor<int32, [4]> input_627_pad_0 = const()[name = tensor<string, []>("input_627_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
-            tensor<int32, []> input_627_groups_0 = const()[name = tensor<string, []>("input_627_groups_0"), val = tensor<int32, []>(1024)];
-            tensor<int32, [2]> input_627_strides_0 = const()[name = tensor<string, []>("input_627_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [2]> input_627_dilations_0 = const()[name = tensor<string, []>("input_627_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [1, 2048, 1, 188]> input_625_cast_fp16 = conv(dilations = input_625_dilations_0, groups = input_625_groups_0, pad = input_625_pad_0, pad_type = input_625_pad_type_0, strides = input_625_strides_0, weight = layers_23_conv_pointwise_conv1_weight_to_fp16, x = input_623_cast_fp16)[name = tensor<string, []>("input_625_cast_fp16")];
+            tensor<int32, []> input_627_split_num_splits_0 = const()[name = tensor<string, []>("input_627_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> input_627_split_axis_0 = const()[name = tensor<string, []>("input_627_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 1, 188]> input_627_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> input_627_split_cast_fp16_1 = split(axis = input_627_split_axis_0, num_splits = input_627_split_num_splits_0, x = input_625_cast_fp16)[name = tensor<string, []>("input_627_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_627_split_1_sigmoid_cast_fp16 = sigmoid(x = input_627_split_cast_fp16_1)[name = tensor<string, []>("input_627_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_627_cast_fp16 = mul(x = input_627_split_cast_fp16_0, y = input_627_split_1_sigmoid_cast_fp16)[name = tensor<string, []>("input_627_cast_fp16")];
+            tensor<string, []> input_629_pad_type_0 = const()[name = tensor<string, []>("input_629_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_629_pad_0 = const()[name = tensor<string, []>("input_629_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
+            tensor<int32, []> input_629_groups_0 = const()[name = tensor<string, []>("input_629_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [2]> input_629_strides_0 = const()[name = tensor<string, []>("input_629_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> input_629_dilations_0 = const()[name = tensor<string, []>("input_629_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<fp16, [1024, 1, 1, 9]> const_314_to_fp16 = const()[name = tensor<string, []>("const_314_to_fp16"), val = tensor<fp16, [1024, 1, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1199626240)))];
             tensor<fp16, [1024]> const_315_to_fp16 = const()[name = tensor<string, []>("const_315_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1199644736)))];
-            tensor<fp16, [1, 1024, 1, 188]> input_629_cast_fp16 = conv(bias = const_315_to_fp16, dilations = input_627_dilations_0, groups = input_627_groups_0, pad = input_627_pad_0, pad_type = input_627_pad_type_0, strides = input_627_strides_0, weight = const_314_to_fp16, x = input_625_cast_fp16)[name = tensor<string, []>("input_629_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> input_631_cast_fp16 = silu(x = input_629_cast_fp16)[name = tensor<string, []>("input_631_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_631_cast_fp16 = conv(bias = const_315_to_fp16, dilations = input_629_dilations_0, groups = input_629_groups_0, pad = input_629_pad_0, pad_type = input_629_pad_type_0, strides = input_629_strides_0, weight = const_314_to_fp16, x = input_627_cast_fp16)[name = tensor<string, []>("input_631_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> input_633_cast_fp16 = silu(x = input_631_cast_fp16)[name = tensor<string, []>("input_633_cast_fp16")];
             tensor<string, []> x_143_pad_type_0 = const()[name = tensor<string, []>("x_143_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_143_strides_0 = const()[name = tensor<string, []>("x_143_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_143_pad_0 = const()[name = tensor<string, []>("x_143_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_143_dilations_0 = const()[name = tensor<string, []>("x_143_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_143_groups_0 = const()[name = tensor<string, []>("x_143_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [1024, 1024, 1, 1]> layers_23_conv_pointwise_conv2_weight_to_fp16 = const()[name = tensor<string, []>("layers_23_conv_pointwise_conv2_weight_to_fp16"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1199646848)))];
-            tensor<fp16, [1, 1024, 1, 188]> x_143_cast_fp16 = conv(dilations = x_143_dilations_0, groups = x_143_groups_0, pad = x_143_pad_0, pad_type = x_143_pad_type_0, strides = x_143_strides_0, weight = layers_23_conv_pointwise_conv2_weight_to_fp16, x = input_631_cast_fp16)[name = tensor<string, []>("x_143_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> x_143_cast_fp16 = conv(dilations = x_143_dilations_0, groups = x_143_groups_0, pad = x_143_pad_0, pad_type = x_143_pad_type_0, strides = x_143_strides_0, weight = layers_23_conv_pointwise_conv2_weight_to_fp16, x = input_633_cast_fp16)[name = tensor<string, []>("x_143_cast_fp16")];
             tensor<fp16, [1, 1024, 1, 188]> inputs_237_cast_fp16 = add(x = inputs_235_cast_fp16, y = x_143_cast_fp16)[name = tensor<string, []>("inputs_237_cast_fp16")];
             tensor<int32, [1]> out_237_axes_0 = const()[name = tensor<string, []>("out_237_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_6849_to_fp16 = const()[name = tensor<string, []>("op_6849_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_237_cast_fp16 = layer_norm(axes = out_237_axes_0, epsilon = var_6849_to_fp16, x = inputs_237_cast_fp16)[name = tensor<string, []>("out_237_cast_fp16")];
-            tensor<fp16, [1024]> input_633_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_633_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1201744064)))];
-            tensor<fp16, [1024]> input_633_beta_0_to_fp16 = const()[name = tensor<string, []>("input_633_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1201746176)))];
-            tensor<fp16, []> input_633_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_633_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> input_633_cast_fp16 = batch_norm(beta = input_633_beta_0_to_fp16, epsilon = input_633_epsilon_0_to_fp16, gamma = input_633_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_237_cast_fp16)[name = tensor<string, []>("input_633_cast_fp16")];
-            tensor<string, []> input_635_pad_type_0 = const()[name = tensor<string, []>("input_635_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> input_635_strides_0 = const()[name = tensor<string, []>("input_635_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> input_635_pad_0 = const()[name = tensor<string, []>("input_635_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> input_635_dilations_0 = const()[name = tensor<string, []>("input_635_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> input_635_groups_0 = const()[name = tensor<string, []>("input_635_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, []> var_6847_to_fp16 = const()[name = tensor<string, []>("op_6847_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_237_cast_fp16 = layer_norm(axes = out_237_axes_0, epsilon = var_6847_to_fp16, x = inputs_237_cast_fp16)[name = tensor<string, []>("out_237_cast_fp16")];
+            tensor<fp16, [1024]> input_635_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_635_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1201744064)))];
+            tensor<fp16, [1024]> input_635_beta_0_to_fp16 = const()[name = tensor<string, []>("input_635_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1201746176)))];
+            tensor<fp16, []> input_635_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_635_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> input_635_cast_fp16 = batch_norm(beta = input_635_beta_0_to_fp16, epsilon = input_635_epsilon_0_to_fp16, gamma = input_635_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_237_cast_fp16)[name = tensor<string, []>("input_635_cast_fp16")];
+            tensor<string, []> input_637_pad_type_0 = const()[name = tensor<string, []>("input_637_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> input_637_strides_0 = const()[name = tensor<string, []>("input_637_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_637_pad_0 = const()[name = tensor<string, []>("input_637_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_637_dilations_0 = const()[name = tensor<string, []>("input_637_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> input_637_groups_0 = const()[name = tensor<string, []>("input_637_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [4096, 1024, 1, 1]> layers_23_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_23_feed_forward2_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1201748288)))];
-            tensor<fp16, [1, 4096, 1, 188]> input_635_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_635_dilations_0, groups = input_635_groups_0, pad = input_635_pad_0, pad_type = input_635_pad_type_0, strides = input_635_strides_0, weight = layers_23_feed_forward2_fc1_weight_to_fp16, x = input_633_cast_fp16)[name = tensor<string, []>("input_635_cast_fp16")];
-            tensor<fp16, [1, 4096, 1, 188]> input_637_cast_fp16 = silu(x = input_635_cast_fp16)[name = tensor<string, []>("input_637_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_637_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_637_dilations_0, groups = input_637_groups_0, pad = input_637_pad_0, pad_type = input_637_pad_type_0, strides = input_637_strides_0, weight = layers_23_feed_forward2_fc1_weight_to_fp16, x = input_635_cast_fp16)[name = tensor<string, []>("input_637_cast_fp16")];
+            tensor<fp16, [1, 4096, 1, 188]> input_639_cast_fp16 = silu(x = input_637_cast_fp16)[name = tensor<string, []>("input_639_cast_fp16")];
             tensor<string, []> x_pad_type_0 = const()[name = tensor<string, []>("x_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [2]> x_strides_0 = const()[name = tensor<string, []>("x_strides_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, [4]> x_pad_0 = const()[name = tensor<string, []>("x_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [2]> x_dilations_0 = const()[name = tensor<string, []>("x_dilations_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> x_groups_0 = const()[name = tensor<string, []>("x_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 4096, 1, 1]> var_6877_weight_0_to_fp16 = const()[name = tensor<string, []>("op_6877_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1210136960)))];
-            tensor<fp16, [1, 1024, 1, 188]> var_6877_cast_fp16 = conv(bias = input_15_mean_0_to_fp16, dilations = x_dilations_0, groups = x_groups_0, pad = x_pad_0, pad_type = x_pad_type_0, strides = x_strides_0, weight = var_6877_weight_0_to_fp16, x = input_637_cast_fp16)[name = tensor<string, []>("op_6877_cast_fp16")];
-            tensor<fp16, [1, 1024, 1, 188]> inputs_cast_fp16 = add(x = inputs_237_cast_fp16, y = var_6877_cast_fp16)[name = tensor<string, []>("inputs_cast_fp16")];
+            tensor<fp16, [1024, 4096, 1, 1]> var_6875_weight_0_to_fp16 = const()[name = tensor<string, []>("op_6875_weight_0_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1210136960)))];
+            tensor<fp16, [1, 1024, 1, 188]> var_6875_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_dilations_0, groups = x_groups_0, pad = x_pad_0, pad_type = x_pad_type_0, strides = x_strides_0, weight = var_6875_weight_0_to_fp16, x = input_639_cast_fp16)[name = tensor<string, []>("op_6875_cast_fp16")];
+            tensor<fp16, [1, 1024, 1, 188]> inputs_cast_fp16 = add(x = inputs_237_cast_fp16, y = var_6875_cast_fp16)[name = tensor<string, []>("inputs_cast_fp16")];
             tensor<int32, [1]> out_239_axes_0 = const()[name = tensor<string, []>("out_239_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, []> var_6887_to_fp16 = const()[name = tensor<string, []>("op_6887_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> out_239_cast_fp16 = layer_norm(axes = out_239_axes_0, epsilon = var_6887_to_fp16, x = inputs_cast_fp16)[name = tensor<string, []>("out_239_cast_fp16")];
+            tensor<fp16, []> var_6885_to_fp16 = const()[name = tensor<string, []>("op_6885_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1024, 1, 188]> out_239_cast_fp16 = layer_norm(axes = out_239_axes_0, epsilon = var_6885_to_fp16, x = inputs_cast_fp16)[name = tensor<string, []>("out_239_cast_fp16")];
             tensor<fp16, [1024]> encoder_output_embeds_type_fp32_gamma_0_to_fp16 = const()[name = tensor<string, []>("encoder_output_embeds_type_fp32_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1218525632)))];
             tensor<fp16, [1024]> encoder_output_embeds_type_fp32_beta_0_to_fp16 = const()[name = tensor<string, []>("encoder_output_embeds_type_fp32_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1218527744)))];
             tensor<fp16, []> encoder_output_embeds_type_fp32_epsilon_0_to_fp16 = const()[name = tensor<string, []>("encoder_output_embeds_type_fp32_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1024, 1, 188]> encoder_output_embeds = batch_norm(beta = encoder_output_embeds_type_fp32_beta_0_to_fp16, epsilon = encoder_output_embeds_type_fp32_epsilon_0_to_fp16, gamma = encoder_output_embeds_type_fp32_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_239_cast_fp16)[name = tensor<string, []>("encoder_output_embeds_type_fp32_cast_fp16")];
-            tensor<string, []> var_6907_pad_type_0 = const()[name = tensor<string, []>("op_6907_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [2]> var_6907_strides_0 = const()[name = tensor<string, []>("op_6907_strides_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, [4]> var_6907_pad_0 = const()[name = tensor<string, []>("op_6907_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
-            tensor<int32, [2]> var_6907_dilations_0 = const()[name = tensor<string, []>("op_6907_dilations_0"), val = tensor<int32, [2]>([1, 1])];
-            tensor<int32, []> var_6907_groups_0 = const()[name = tensor<string, []>("op_6907_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 1, 188]> encoder_output_embeds = batch_norm(beta = encoder_output_embeds_type_fp32_beta_0_to_fp16, epsilon = encoder_output_embeds_type_fp32_epsilon_0_to_fp16, gamma = encoder_output_embeds_type_fp32_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_239_cast_fp16)[name = tensor<string, []>("encoder_output_embeds_type_fp32_cast_fp16")];
+            tensor<string, []> var_6905_pad_type_0 = const()[name = tensor<string, []>("op_6905_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> var_6905_strides_0 = const()[name = tensor<string, []>("op_6905_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> var_6905_pad_0 = const()[name = tensor<string, []>("op_6905_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> var_6905_dilations_0 = const()[name = tensor<string, []>("op_6905_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> var_6905_groups_0 = const()[name = tensor<string, []>("op_6905_groups_0"), val = tensor<int32, []>(1)];
             tensor<fp16, [640, 1024, 1, 1]> joint_projection_weight_to_fp16 = const()[name = tensor<string, []>("joint_projection_weight_to_fp16"), val = tensor<fp16, [640, 1024, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1218529856)))];
             tensor<fp16, [640]> joint_projection_bias_to_fp16 = const()[name = tensor<string, []>("joint_projection_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1219840640)))];
-            tensor<fp16, [1, 640, 1, 188]> joint_projected_encoder_output_embeds = conv(bias = joint_projection_bias_to_fp16, dilations = var_6907_dilations_0, groups = var_6907_groups_0, pad = var_6907_pad_0, pad_type = var_6907_pad_type_0, strides = var_6907_strides_0, weight = joint_projection_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("op_6907_cast_fp16")];
+            tensor<fp16, [1, 640, 1, 188]> joint_projected_encoder_output_embeds = conv(bias = joint_projection_bias_to_fp16, dilations = var_6905_dilations_0, groups = var_6905_groups_0, pad = var_6905_pad_0, pad_type = var_6905_pad_type_0, strides = var_6905_strides_0, weight = joint_projection_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("op_6905_cast_fp16")];
         } -> (encoder_output_embeds, joint_projected_encoder_output_embeds);
 }
\ No newline at end of file