diff --git "a/distil-whisper_distil-large-v3/AudioEncoder.mlmodelc/model.mil" "b/distil-whisper_distil-large-v3/AudioEncoder.mlmodelc/model.mil" new file mode 100644--- /dev/null +++ "b/distil-whisper_distil-large-v3/AudioEncoder.mlmodelc/model.mil" @@ -0,0 +1,2737 @@ +program(1.3) +[buildInfo = dict({{"coremlc-component-MIL", "3401.3.1"}, {"coremlc-version", "3401.4.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.0"}})] +{ + func main(tensor melspectrogram_features) { + string var_110_pad_type_0 = const()[name = string("op_110_pad_type_0"), val = string("custom")]; + tensor var_110_pad_0 = const()[name = string("op_110_pad_0"), val = tensor([0, 0, 1, 1])]; + tensor var_110_strides_0 = const()[name = string("op_110_strides_0"), val = tensor([1, 1])]; + tensor var_110_dilations_0 = const()[name = string("op_110_dilations_0"), val = tensor([1, 1])]; + int32 var_110_groups_0 = const()[name = string("op_110_groups_0"), val = int32(1)]; + tensor var_85_to_fp16 = const()[name = string("op_85_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; + tensor var_91_to_fp16 = const()[name = string("op_91_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(983168)))]; + tensor var_110_cast_fp16 = conv(bias = var_91_to_fp16, dilations = var_110_dilations_0, groups = var_110_groups_0, pad = var_110_pad_0, pad_type = var_110_pad_type_0, strides = var_110_strides_0, weight = var_85_to_fp16, x = melspectrogram_features)[name = string("op_110_cast_fp16")]; + string hidden_states_1_mode_0 = const()[name = string("hidden_states_1_mode_0"), val = string("EXACT")]; + tensor hidden_states_1_cast_fp16 = gelu(mode = hidden_states_1_mode_0, x = var_110_cast_fp16)[name = string("hidden_states_1_cast_fp16")]; + string var_150_pad_type_0 = const()[name = string("op_150_pad_type_0"), val = string("custom")]; + tensor var_150_pad_0 = const()[name = string("op_150_pad_0"), val = tensor([0, 0, 1, 1])]; + tensor var_150_strides_0 = const()[name = string("op_150_strides_0"), val = tensor([2, 2])]; + tensor var_150_dilations_0 = const()[name = string("op_150_dilations_0"), val = tensor([1, 1])]; + int32 var_150_groups_0 = const()[name = string("op_150_groups_0"), val = int32(1)]; + tensor var_125_to_fp16 = const()[name = string("op_125_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(985792)))]; + tensor var_131_to_fp16 = const()[name = string("op_131_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10816256)))]; + tensor var_150_cast_fp16 = conv(bias = var_131_to_fp16, dilations = var_150_dilations_0, groups = var_150_groups_0, pad = var_150_pad_0, pad_type = var_150_pad_type_0, strides = var_150_strides_0, weight = var_125_to_fp16, x = hidden_states_1_cast_fp16)[name = string("op_150_cast_fp16")]; + string hidden_states_3_mode_0 = const()[name = string("hidden_states_3_mode_0"), val = string("EXACT")]; + tensor hidden_states_3_cast_fp16 = gelu(mode = hidden_states_3_mode_0, x = var_150_cast_fp16)[name = string("hidden_states_3_cast_fp16")]; + tensor var_168_to_fp16 = const()[name = string("op_168_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10818880)))]; + tensor inputs_1_cast_fp16 = add(x = hidden_states_3_cast_fp16, y = var_168_to_fp16)[name = string("inputs_1_cast_fp16")]; + int32 var_182 = const()[name = string("op_182"), val = int32(3)]; + tensor out_1_axes_0 = const()[name = string("out_1_axes_0"), val = tensor([1])]; + fp16 var_201_to_fp16 = const()[name = string("op_201_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_201_to_fp16, x = inputs_1_cast_fp16)[name = string("out_1_cast_fp16")]; + tensor obj_1_mean_0_to_fp16 = const()[name = string("obj_1_mean_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14658944)))]; + tensor obj_1_variance_0_to_fp16 = const()[name = string("obj_1_variance_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14661568)))]; + tensor obj_1_gamma_0_to_fp16 = const()[name = string("obj_1_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14664192)))]; + tensor obj_1_beta_0_to_fp16 = const()[name = string("obj_1_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14666816)))]; + fp16 obj_1_epsilon_0_to_fp16 = const()[name = string("obj_1_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor 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 = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_1_cast_fp16)[name = string("obj_1_cast_fp16")]; + string query_1_pad_type_0 = const()[name = string("query_1_pad_type_0"), val = string("valid")]; + tensor query_1_strides_0 = const()[name = string("query_1_strides_0"), val = tensor([1, 1])]; + tensor query_1_pad_0 = const()[name = string("query_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_1_dilations_0 = const()[name = string("query_1_dilations_0"), val = tensor([1, 1])]; + int32 query_1_groups_0 = const()[name = string("query_1_groups_0"), val = int32(1)]; + tensor layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14669440)))]; + tensor layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17946304)))]; + tensor query_1_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_bias_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 = string("query_1_cast_fp16")]; + string key_1_pad_type_0 = const()[name = string("key_1_pad_type_0"), val = string("valid")]; + tensor key_1_strides_0 = const()[name = string("key_1_strides_0"), val = tensor([1, 1])]; + tensor key_1_pad_0 = const()[name = string("key_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_1_dilations_0 = const()[name = string("key_1_dilations_0"), val = tensor([1, 1])]; + int32 key_1_groups_0 = const()[name = string("key_1_groups_0"), val = int32(1)]; + tensor layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17948928)))]; + tensor 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 = string("key_1_cast_fp16")]; + string value_1_pad_type_0 = const()[name = string("value_1_pad_type_0"), val = string("valid")]; + tensor value_1_strides_0 = const()[name = string("value_1_strides_0"), val = tensor([1, 1])]; + tensor value_1_pad_0 = const()[name = string("value_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_1_dilations_0 = const()[name = string("value_1_dilations_0"), val = tensor([1, 1])]; + int32 value_1_groups_0 = const()[name = string("value_1_groups_0"), val = int32(1)]; + tensor layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21225792)))]; + tensor layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24502656)))]; + tensor value_1_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_bias_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 = string("value_1_cast_fp16")]; + tensor var_236 = const()[name = string("op_236"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_1_cast_fp16 = reshape(shape = var_236, x = query_1_cast_fp16)[name = string("mh_q_1_cast_fp16")]; + fp16 var_238_to_fp16 = const()[name = string("op_238_to_fp16"), val = fp16(0x1p-3)]; + tensor var_239_cast_fp16 = mul(x = mh_q_1_cast_fp16, y = var_238_to_fp16)[name = string("op_239_cast_fp16")]; + tensor var_240 = const()[name = string("op_240"), val = tensor([1, 20, 64, -1])]; + tensor var_241_cast_fp16 = reshape(shape = var_240, x = key_1_cast_fp16)[name = string("op_241_cast_fp16")]; + bool mh_w_1_transpose_x_0 = const()[name = string("mh_w_1_transpose_x_0"), val = bool(true)]; + bool mh_w_1_transpose_y_0 = const()[name = string("mh_w_1_transpose_y_0"), val = bool(false)]; + tensor mh_w_1_cast_fp16 = matmul(transpose_x = mh_w_1_transpose_x_0, transpose_y = mh_w_1_transpose_y_0, x = var_239_cast_fp16, y = var_241_cast_fp16)[name = string("mh_w_1_cast_fp16")]; + tensor var_244_cast_fp16 = softmax(axis = var_182, x = mh_w_1_cast_fp16)[name = string("op_244_cast_fp16")]; + tensor var_245 = const()[name = string("op_245"), val = tensor([1, 20, 64, -1])]; + tensor var_246_cast_fp16 = reshape(shape = var_245, x = value_1_cast_fp16)[name = string("op_246_cast_fp16")]; + bool attn_1_transpose_x_0 = const()[name = string("attn_1_transpose_x_0"), val = bool(false)]; + bool attn_1_transpose_y_0 = const()[name = string("attn_1_transpose_y_0"), val = bool(true)]; + tensor attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = var_246_cast_fp16, y = var_244_cast_fp16)[name = string("attn_1_cast_fp16")]; + tensor var_249 = const()[name = string("op_249"), val = tensor([1, 1280, 1, -1])]; + tensor input_1_cast_fp16 = reshape(shape = var_249, x = attn_1_cast_fp16)[name = string("input_1_cast_fp16")]; + string obj_3_pad_type_0 = const()[name = string("obj_3_pad_type_0"), val = string("valid")]; + tensor obj_3_strides_0 = const()[name = string("obj_3_strides_0"), val = tensor([1, 1])]; + tensor obj_3_pad_0 = const()[name = string("obj_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_3_dilations_0 = const()[name = string("obj_3_dilations_0"), val = tensor([1, 1])]; + int32 obj_3_groups_0 = const()[name = string("obj_3_groups_0"), val = int32(1)]; + tensor layers_0_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24505280)))]; + tensor layers_0_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_0_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27782144)))]; + tensor obj_3_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_bias_to_fp16, dilations = obj_3_dilations_0, groups = obj_3_groups_0, pad = obj_3_pad_0, pad_type = obj_3_pad_type_0, strides = obj_3_strides_0, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = input_1_cast_fp16)[name = string("obj_3_cast_fp16")]; + tensor inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = obj_3_cast_fp16)[name = string("inputs_3_cast_fp16")]; + tensor out_3_axes_0 = const()[name = string("out_3_axes_0"), val = tensor([1])]; + fp16 var_267_to_fp16 = const()[name = string("op_267_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_267_to_fp16, x = inputs_3_cast_fp16)[name = string("out_3_cast_fp16")]; + tensor input_3_gamma_0_to_fp16 = const()[name = string("input_3_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27784768)))]; + tensor input_3_beta_0_to_fp16 = const()[name = string("input_3_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27787392)))]; + fp16 input_3_epsilon_0_to_fp16 = const()[name = string("input_3_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor input_3_cast_fp16 = batch_norm(beta = input_3_beta_0_to_fp16, epsilon = input_3_epsilon_0_to_fp16, gamma = input_3_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_3_cast_fp16)[name = string("input_3_cast_fp16")]; + string input_5_pad_type_0 = const()[name = string("input_5_pad_type_0"), val = string("valid")]; + tensor input_5_strides_0 = const()[name = string("input_5_strides_0"), val = tensor([1, 1])]; + tensor input_5_pad_0 = const()[name = string("input_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_5_dilations_0 = const()[name = string("input_5_dilations_0"), val = tensor([1, 1])]; + int32 input_5_groups_0 = const()[name = string("input_5_groups_0"), val = int32(1)]; + tensor layers_0_fc1_weight_to_fp16 = const()[name = string("layers_0_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27790016)))]; + tensor layers_0_fc1_bias_to_fp16 = const()[name = string("layers_0_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40897280)))]; + tensor input_5_cast_fp16 = conv(bias = layers_0_fc1_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 = layers_0_fc1_weight_to_fp16, x = input_3_cast_fp16)[name = string("input_5_cast_fp16")]; + string input_7_mode_0 = const()[name = string("input_7_mode_0"), val = string("EXACT")]; + tensor input_7_cast_fp16 = gelu(mode = input_7_mode_0, x = input_5_cast_fp16)[name = string("input_7_cast_fp16")]; + string hidden_states_5_pad_type_0 = const()[name = string("hidden_states_5_pad_type_0"), val = string("valid")]; + tensor hidden_states_5_strides_0 = const()[name = string("hidden_states_5_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_5_pad_0 = const()[name = string("hidden_states_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_5_dilations_0 = const()[name = string("hidden_states_5_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_5_groups_0 = const()[name = string("hidden_states_5_groups_0"), val = int32(1)]; + tensor layers_0_fc2_weight_to_fp16 = const()[name = string("layers_0_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40907584)))]; + tensor layers_0_fc2_bias_to_fp16 = const()[name = string("layers_0_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54014848)))]; + tensor hidden_states_5_cast_fp16 = conv(bias = layers_0_fc2_bias_to_fp16, dilations = hidden_states_5_dilations_0, groups = hidden_states_5_groups_0, pad = hidden_states_5_pad_0, pad_type = hidden_states_5_pad_type_0, strides = hidden_states_5_strides_0, weight = layers_0_fc2_weight_to_fp16, x = input_7_cast_fp16)[name = string("hidden_states_5_cast_fp16")]; + tensor inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = hidden_states_5_cast_fp16)[name = string("inputs_5_cast_fp16")]; + int32 var_300 = const()[name = string("op_300"), val = int32(3)]; + tensor out_5_axes_0 = const()[name = string("out_5_axes_0"), val = tensor([1])]; + fp16 var_319_to_fp16 = const()[name = string("op_319_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_319_to_fp16, x = inputs_5_cast_fp16)[name = string("out_5_cast_fp16")]; + tensor obj_5_gamma_0_to_fp16 = const()[name = string("obj_5_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54017472)))]; + tensor obj_5_beta_0_to_fp16 = const()[name = string("obj_5_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54020096)))]; + fp16 obj_5_epsilon_0_to_fp16 = const()[name = string("obj_5_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor obj_5_cast_fp16 = batch_norm(beta = obj_5_beta_0_to_fp16, epsilon = obj_5_epsilon_0_to_fp16, gamma = obj_5_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_5_cast_fp16)[name = string("obj_5_cast_fp16")]; + string query_3_pad_type_0 = const()[name = string("query_3_pad_type_0"), val = string("valid")]; + tensor query_3_strides_0 = const()[name = string("query_3_strides_0"), val = tensor([1, 1])]; + tensor query_3_pad_0 = const()[name = string("query_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_3_dilations_0 = const()[name = string("query_3_dilations_0"), val = tensor([1, 1])]; + int32 query_3_groups_0 = const()[name = string("query_3_groups_0"), val = int32(1)]; + tensor layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54022720)))]; + tensor layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57299584)))]; + tensor query_3_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_bias_to_fp16, dilations = query_3_dilations_0, groups = query_3_groups_0, pad = query_3_pad_0, pad_type = query_3_pad_type_0, strides = query_3_strides_0, weight = layers_1_self_attn_q_proj_weight_to_fp16, x = obj_5_cast_fp16)[name = string("query_3_cast_fp16")]; + string key_3_pad_type_0 = const()[name = string("key_3_pad_type_0"), val = string("valid")]; + tensor key_3_strides_0 = const()[name = string("key_3_strides_0"), val = tensor([1, 1])]; + tensor key_3_pad_0 = const()[name = string("key_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_3_dilations_0 = const()[name = string("key_3_dilations_0"), val = tensor([1, 1])]; + int32 key_3_groups_0 = const()[name = string("key_3_groups_0"), val = int32(1)]; + tensor layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57302208)))]; + tensor key_3_cast_fp16 = conv(dilations = key_3_dilations_0, groups = key_3_groups_0, pad = key_3_pad_0, pad_type = key_3_pad_type_0, strides = key_3_strides_0, weight = layers_1_self_attn_k_proj_weight_to_fp16, x = obj_5_cast_fp16)[name = string("key_3_cast_fp16")]; + string value_3_pad_type_0 = const()[name = string("value_3_pad_type_0"), val = string("valid")]; + tensor value_3_strides_0 = const()[name = string("value_3_strides_0"), val = tensor([1, 1])]; + tensor value_3_pad_0 = const()[name = string("value_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_3_dilations_0 = const()[name = string("value_3_dilations_0"), val = tensor([1, 1])]; + int32 value_3_groups_0 = const()[name = string("value_3_groups_0"), val = int32(1)]; + tensor layers_1_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60579072)))]; + tensor layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63855936)))]; + tensor value_3_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_bias_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_5_cast_fp16)[name = string("value_3_cast_fp16")]; + tensor var_354 = const()[name = string("op_354"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_3_cast_fp16 = reshape(shape = var_354, x = query_3_cast_fp16)[name = string("mh_q_3_cast_fp16")]; + fp16 var_356_to_fp16 = const()[name = string("op_356_to_fp16"), val = fp16(0x1p-3)]; + tensor var_357_cast_fp16 = mul(x = mh_q_3_cast_fp16, y = var_356_to_fp16)[name = string("op_357_cast_fp16")]; + tensor var_358 = const()[name = string("op_358"), val = tensor([1, 20, 64, -1])]; + tensor var_359_cast_fp16 = reshape(shape = var_358, x = key_3_cast_fp16)[name = string("op_359_cast_fp16")]; + bool mh_w_3_transpose_x_0 = const()[name = string("mh_w_3_transpose_x_0"), val = bool(true)]; + bool mh_w_3_transpose_y_0 = const()[name = string("mh_w_3_transpose_y_0"), val = bool(false)]; + tensor mh_w_3_cast_fp16 = matmul(transpose_x = mh_w_3_transpose_x_0, transpose_y = mh_w_3_transpose_y_0, x = var_357_cast_fp16, y = var_359_cast_fp16)[name = string("mh_w_3_cast_fp16")]; + tensor var_362_cast_fp16 = softmax(axis = var_300, x = mh_w_3_cast_fp16)[name = string("op_362_cast_fp16")]; + tensor var_363 = const()[name = string("op_363"), val = tensor([1, 20, 64, -1])]; + tensor var_364_cast_fp16 = reshape(shape = var_363, x = value_3_cast_fp16)[name = string("op_364_cast_fp16")]; + bool attn_3_transpose_x_0 = const()[name = string("attn_3_transpose_x_0"), val = bool(false)]; + bool attn_3_transpose_y_0 = const()[name = string("attn_3_transpose_y_0"), val = bool(true)]; + tensor attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = var_364_cast_fp16, y = var_362_cast_fp16)[name = string("attn_3_cast_fp16")]; + tensor var_367 = const()[name = string("op_367"), val = tensor([1, 1280, 1, -1])]; + tensor input_9_cast_fp16 = reshape(shape = var_367, x = attn_3_cast_fp16)[name = string("input_9_cast_fp16")]; + string obj_7_pad_type_0 = const()[name = string("obj_7_pad_type_0"), val = string("valid")]; + tensor obj_7_strides_0 = const()[name = string("obj_7_strides_0"), val = tensor([1, 1])]; + tensor obj_7_pad_0 = const()[name = string("obj_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_7_dilations_0 = const()[name = string("obj_7_dilations_0"), val = tensor([1, 1])]; + int32 obj_7_groups_0 = const()[name = string("obj_7_groups_0"), val = int32(1)]; + tensor layers_1_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63858560)))]; + tensor layers_1_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_1_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67135424)))]; + tensor obj_7_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_bias_to_fp16, dilations = obj_7_dilations_0, groups = obj_7_groups_0, pad = obj_7_pad_0, pad_type = obj_7_pad_type_0, strides = obj_7_strides_0, weight = layers_1_self_attn_o_proj_weight_to_fp16, x = input_9_cast_fp16)[name = string("obj_7_cast_fp16")]; + tensor inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = obj_7_cast_fp16)[name = string("inputs_7_cast_fp16")]; + tensor out_7_axes_0 = const()[name = string("out_7_axes_0"), val = tensor([1])]; + fp16 var_385_to_fp16 = const()[name = string("op_385_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_385_to_fp16, x = inputs_7_cast_fp16)[name = string("out_7_cast_fp16")]; + tensor input_11_gamma_0_to_fp16 = const()[name = string("input_11_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67138048)))]; + tensor input_11_beta_0_to_fp16 = const()[name = string("input_11_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67140672)))]; + fp16 input_11_epsilon_0_to_fp16 = const()[name = string("input_11_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor input_11_cast_fp16 = batch_norm(beta = input_11_beta_0_to_fp16, epsilon = input_11_epsilon_0_to_fp16, gamma = input_11_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_7_cast_fp16)[name = string("input_11_cast_fp16")]; + string input_13_pad_type_0 = const()[name = string("input_13_pad_type_0"), val = string("valid")]; + tensor input_13_strides_0 = const()[name = string("input_13_strides_0"), val = tensor([1, 1])]; + tensor input_13_pad_0 = const()[name = string("input_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_13_dilations_0 = const()[name = string("input_13_dilations_0"), val = tensor([1, 1])]; + int32 input_13_groups_0 = const()[name = string("input_13_groups_0"), val = int32(1)]; + tensor layers_1_fc1_weight_to_fp16 = const()[name = string("layers_1_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67143296)))]; + tensor layers_1_fc1_bias_to_fp16 = const()[name = string("layers_1_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80250560)))]; + tensor input_13_cast_fp16 = conv(bias = layers_1_fc1_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 = layers_1_fc1_weight_to_fp16, x = input_11_cast_fp16)[name = string("input_13_cast_fp16")]; + string input_15_mode_0 = const()[name = string("input_15_mode_0"), val = string("EXACT")]; + tensor input_15_cast_fp16 = gelu(mode = input_15_mode_0, x = input_13_cast_fp16)[name = string("input_15_cast_fp16")]; + string hidden_states_7_pad_type_0 = const()[name = string("hidden_states_7_pad_type_0"), val = string("valid")]; + tensor hidden_states_7_strides_0 = const()[name = string("hidden_states_7_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_7_pad_0 = const()[name = string("hidden_states_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_7_dilations_0 = const()[name = string("hidden_states_7_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_7_groups_0 = const()[name = string("hidden_states_7_groups_0"), val = int32(1)]; + tensor layers_1_fc2_weight_to_fp16 = const()[name = string("layers_1_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80260864)))]; + tensor layers_1_fc2_bias_to_fp16 = const()[name = string("layers_1_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93368128)))]; + tensor hidden_states_7_cast_fp16 = conv(bias = layers_1_fc2_bias_to_fp16, dilations = hidden_states_7_dilations_0, groups = hidden_states_7_groups_0, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = hidden_states_7_strides_0, weight = layers_1_fc2_weight_to_fp16, x = input_15_cast_fp16)[name = string("hidden_states_7_cast_fp16")]; + tensor inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = hidden_states_7_cast_fp16)[name = string("inputs_9_cast_fp16")]; + int32 var_418 = const()[name = string("op_418"), val = int32(3)]; + tensor out_9_axes_0 = const()[name = string("out_9_axes_0"), val = tensor([1])]; + fp16 var_437_to_fp16 = const()[name = string("op_437_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_437_to_fp16, x = inputs_9_cast_fp16)[name = string("out_9_cast_fp16")]; + tensor obj_9_gamma_0_to_fp16 = const()[name = string("obj_9_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93370752)))]; + tensor obj_9_beta_0_to_fp16 = const()[name = string("obj_9_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93373376)))]; + fp16 obj_9_epsilon_0_to_fp16 = const()[name = string("obj_9_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor obj_9_cast_fp16 = batch_norm(beta = obj_9_beta_0_to_fp16, epsilon = obj_9_epsilon_0_to_fp16, gamma = obj_9_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_9_cast_fp16)[name = string("obj_9_cast_fp16")]; + string query_5_pad_type_0 = const()[name = string("query_5_pad_type_0"), val = string("valid")]; + tensor query_5_strides_0 = const()[name = string("query_5_strides_0"), val = tensor([1, 1])]; + tensor query_5_pad_0 = const()[name = string("query_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_5_dilations_0 = const()[name = string("query_5_dilations_0"), val = tensor([1, 1])]; + int32 query_5_groups_0 = const()[name = string("query_5_groups_0"), val = int32(1)]; + tensor layers_2_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93376000)))]; + tensor layers_2_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_2_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96652864)))]; + tensor query_5_cast_fp16 = conv(bias = layers_2_self_attn_q_proj_bias_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_2_self_attn_q_proj_weight_to_fp16, x = obj_9_cast_fp16)[name = string("query_5_cast_fp16")]; + string key_5_pad_type_0 = const()[name = string("key_5_pad_type_0"), val = string("valid")]; + tensor key_5_strides_0 = const()[name = string("key_5_strides_0"), val = tensor([1, 1])]; + tensor key_5_pad_0 = const()[name = string("key_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_5_dilations_0 = const()[name = string("key_5_dilations_0"), val = tensor([1, 1])]; + int32 key_5_groups_0 = const()[name = string("key_5_groups_0"), val = int32(1)]; + tensor layers_2_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96655488)))]; + tensor key_5_cast_fp16 = conv(dilations = key_5_dilations_0, groups = key_5_groups_0, pad = key_5_pad_0, pad_type = key_5_pad_type_0, strides = key_5_strides_0, weight = layers_2_self_attn_k_proj_weight_to_fp16, x = obj_9_cast_fp16)[name = string("key_5_cast_fp16")]; + string value_5_pad_type_0 = const()[name = string("value_5_pad_type_0"), val = string("valid")]; + tensor value_5_strides_0 = const()[name = string("value_5_strides_0"), val = tensor([1, 1])]; + tensor value_5_pad_0 = const()[name = string("value_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_5_dilations_0 = const()[name = string("value_5_dilations_0"), val = tensor([1, 1])]; + int32 value_5_groups_0 = const()[name = string("value_5_groups_0"), val = int32(1)]; + tensor layers_2_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99932352)))]; + tensor layers_2_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_2_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103209216)))]; + tensor value_5_cast_fp16 = conv(bias = layers_2_self_attn_v_proj_bias_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_9_cast_fp16)[name = string("value_5_cast_fp16")]; + tensor var_472 = const()[name = string("op_472"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_5_cast_fp16 = reshape(shape = var_472, x = query_5_cast_fp16)[name = string("mh_q_5_cast_fp16")]; + fp16 var_474_to_fp16 = const()[name = string("op_474_to_fp16"), val = fp16(0x1p-3)]; + tensor var_475_cast_fp16 = mul(x = mh_q_5_cast_fp16, y = var_474_to_fp16)[name = string("op_475_cast_fp16")]; + tensor var_476 = const()[name = string("op_476"), val = tensor([1, 20, 64, -1])]; + tensor var_477_cast_fp16 = reshape(shape = var_476, x = key_5_cast_fp16)[name = string("op_477_cast_fp16")]; + bool mh_w_5_transpose_x_0 = const()[name = string("mh_w_5_transpose_x_0"), val = bool(true)]; + bool mh_w_5_transpose_y_0 = const()[name = string("mh_w_5_transpose_y_0"), val = bool(false)]; + tensor mh_w_5_cast_fp16 = matmul(transpose_x = mh_w_5_transpose_x_0, transpose_y = mh_w_5_transpose_y_0, x = var_475_cast_fp16, y = var_477_cast_fp16)[name = string("mh_w_5_cast_fp16")]; + tensor var_480_cast_fp16 = softmax(axis = var_418, x = mh_w_5_cast_fp16)[name = string("op_480_cast_fp16")]; + tensor var_481 = const()[name = string("op_481"), val = tensor([1, 20, 64, -1])]; + tensor var_482_cast_fp16 = reshape(shape = var_481, x = value_5_cast_fp16)[name = string("op_482_cast_fp16")]; + bool attn_5_transpose_x_0 = const()[name = string("attn_5_transpose_x_0"), val = bool(false)]; + bool attn_5_transpose_y_0 = const()[name = string("attn_5_transpose_y_0"), val = bool(true)]; + tensor attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = var_482_cast_fp16, y = var_480_cast_fp16)[name = string("attn_5_cast_fp16")]; + tensor var_485 = const()[name = string("op_485"), val = tensor([1, 1280, 1, -1])]; + tensor input_17_cast_fp16 = reshape(shape = var_485, x = attn_5_cast_fp16)[name = string("input_17_cast_fp16")]; + string obj_11_pad_type_0 = const()[name = string("obj_11_pad_type_0"), val = string("valid")]; + tensor obj_11_strides_0 = const()[name = string("obj_11_strides_0"), val = tensor([1, 1])]; + tensor obj_11_pad_0 = const()[name = string("obj_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_11_dilations_0 = const()[name = string("obj_11_dilations_0"), val = tensor([1, 1])]; + int32 obj_11_groups_0 = const()[name = string("obj_11_groups_0"), val = int32(1)]; + tensor layers_2_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103211840)))]; + tensor layers_2_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_2_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106488704)))]; + tensor obj_11_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_bias_to_fp16, dilations = obj_11_dilations_0, groups = obj_11_groups_0, pad = obj_11_pad_0, pad_type = obj_11_pad_type_0, strides = obj_11_strides_0, weight = layers_2_self_attn_o_proj_weight_to_fp16, x = input_17_cast_fp16)[name = string("obj_11_cast_fp16")]; + tensor inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = obj_11_cast_fp16)[name = string("inputs_11_cast_fp16")]; + tensor out_11_axes_0 = const()[name = string("out_11_axes_0"), val = tensor([1])]; + fp16 var_503_to_fp16 = const()[name = string("op_503_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_503_to_fp16, x = inputs_11_cast_fp16)[name = string("out_11_cast_fp16")]; + tensor input_19_gamma_0_to_fp16 = const()[name = string("input_19_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106491328)))]; + tensor input_19_beta_0_to_fp16 = const()[name = string("input_19_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106493952)))]; + fp16 input_19_epsilon_0_to_fp16 = const()[name = string("input_19_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor input_19_cast_fp16 = batch_norm(beta = input_19_beta_0_to_fp16, epsilon = input_19_epsilon_0_to_fp16, gamma = input_19_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_11_cast_fp16)[name = string("input_19_cast_fp16")]; + string input_21_pad_type_0 = const()[name = string("input_21_pad_type_0"), val = string("valid")]; + tensor input_21_strides_0 = const()[name = string("input_21_strides_0"), val = tensor([1, 1])]; + tensor input_21_pad_0 = const()[name = string("input_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_21_dilations_0 = const()[name = string("input_21_dilations_0"), val = tensor([1, 1])]; + int32 input_21_groups_0 = const()[name = string("input_21_groups_0"), val = int32(1)]; + tensor layers_2_fc1_weight_to_fp16 = const()[name = string("layers_2_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106496576)))]; + tensor layers_2_fc1_bias_to_fp16 = const()[name = string("layers_2_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119603840)))]; + tensor input_21_cast_fp16 = conv(bias = layers_2_fc1_bias_to_fp16, dilations = input_21_dilations_0, groups = input_21_groups_0, pad = input_21_pad_0, pad_type = input_21_pad_type_0, strides = input_21_strides_0, weight = layers_2_fc1_weight_to_fp16, x = input_19_cast_fp16)[name = string("input_21_cast_fp16")]; + string input_23_mode_0 = const()[name = string("input_23_mode_0"), val = string("EXACT")]; + tensor input_23_cast_fp16 = gelu(mode = input_23_mode_0, x = input_21_cast_fp16)[name = string("input_23_cast_fp16")]; + string hidden_states_9_pad_type_0 = const()[name = string("hidden_states_9_pad_type_0"), val = string("valid")]; + tensor hidden_states_9_strides_0 = const()[name = string("hidden_states_9_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_9_pad_0 = const()[name = string("hidden_states_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_9_dilations_0 = const()[name = string("hidden_states_9_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_9_groups_0 = const()[name = string("hidden_states_9_groups_0"), val = int32(1)]; + tensor layers_2_fc2_weight_to_fp16 = const()[name = string("layers_2_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119614144)))]; + tensor layers_2_fc2_bias_to_fp16 = const()[name = string("layers_2_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132721408)))]; + tensor hidden_states_9_cast_fp16 = conv(bias = layers_2_fc2_bias_to_fp16, dilations = hidden_states_9_dilations_0, groups = hidden_states_9_groups_0, pad = hidden_states_9_pad_0, pad_type = hidden_states_9_pad_type_0, strides = hidden_states_9_strides_0, weight = layers_2_fc2_weight_to_fp16, x = input_23_cast_fp16)[name = string("hidden_states_9_cast_fp16")]; + tensor inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = hidden_states_9_cast_fp16)[name = string("inputs_13_cast_fp16")]; + int32 var_536 = const()[name = string("op_536"), val = int32(3)]; + tensor out_13_axes_0 = const()[name = string("out_13_axes_0"), val = tensor([1])]; + fp16 var_555_to_fp16 = const()[name = string("op_555_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_555_to_fp16, x = inputs_13_cast_fp16)[name = string("out_13_cast_fp16")]; + tensor obj_13_gamma_0_to_fp16 = const()[name = string("obj_13_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132724032)))]; + tensor obj_13_beta_0_to_fp16 = const()[name = string("obj_13_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132726656)))]; + fp16 obj_13_epsilon_0_to_fp16 = const()[name = string("obj_13_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor obj_13_cast_fp16 = batch_norm(beta = obj_13_beta_0_to_fp16, epsilon = obj_13_epsilon_0_to_fp16, gamma = obj_13_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_13_cast_fp16)[name = string("obj_13_cast_fp16")]; + string query_7_pad_type_0 = const()[name = string("query_7_pad_type_0"), val = string("valid")]; + tensor query_7_strides_0 = const()[name = string("query_7_strides_0"), val = tensor([1, 1])]; + tensor query_7_pad_0 = const()[name = string("query_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_7_dilations_0 = const()[name = string("query_7_dilations_0"), val = tensor([1, 1])]; + int32 query_7_groups_0 = const()[name = string("query_7_groups_0"), val = int32(1)]; + tensor layers_3_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132729280)))]; + tensor layers_3_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_3_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136006144)))]; + tensor query_7_cast_fp16 = conv(bias = layers_3_self_attn_q_proj_bias_to_fp16, dilations = query_7_dilations_0, groups = query_7_groups_0, pad = query_7_pad_0, pad_type = query_7_pad_type_0, strides = query_7_strides_0, weight = layers_3_self_attn_q_proj_weight_to_fp16, x = obj_13_cast_fp16)[name = string("query_7_cast_fp16")]; + string key_7_pad_type_0 = const()[name = string("key_7_pad_type_0"), val = string("valid")]; + tensor key_7_strides_0 = const()[name = string("key_7_strides_0"), val = tensor([1, 1])]; + tensor key_7_pad_0 = const()[name = string("key_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_7_dilations_0 = const()[name = string("key_7_dilations_0"), val = tensor([1, 1])]; + int32 key_7_groups_0 = const()[name = string("key_7_groups_0"), val = int32(1)]; + tensor layers_3_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136008768)))]; + tensor key_7_cast_fp16 = conv(dilations = key_7_dilations_0, groups = key_7_groups_0, pad = key_7_pad_0, pad_type = key_7_pad_type_0, strides = key_7_strides_0, weight = layers_3_self_attn_k_proj_weight_to_fp16, x = obj_13_cast_fp16)[name = string("key_7_cast_fp16")]; + string value_7_pad_type_0 = const()[name = string("value_7_pad_type_0"), val = string("valid")]; + tensor value_7_strides_0 = const()[name = string("value_7_strides_0"), val = tensor([1, 1])]; + tensor value_7_pad_0 = const()[name = string("value_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_7_dilations_0 = const()[name = string("value_7_dilations_0"), val = tensor([1, 1])]; + int32 value_7_groups_0 = const()[name = string("value_7_groups_0"), val = int32(1)]; + tensor layers_3_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139285632)))]; + tensor layers_3_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_3_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142562496)))]; + tensor value_7_cast_fp16 = conv(bias = layers_3_self_attn_v_proj_bias_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_13_cast_fp16)[name = string("value_7_cast_fp16")]; + tensor var_590 = const()[name = string("op_590"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_7_cast_fp16 = reshape(shape = var_590, x = query_7_cast_fp16)[name = string("mh_q_7_cast_fp16")]; + fp16 var_592_to_fp16 = const()[name = string("op_592_to_fp16"), val = fp16(0x1p-3)]; + tensor var_593_cast_fp16 = mul(x = mh_q_7_cast_fp16, y = var_592_to_fp16)[name = string("op_593_cast_fp16")]; + tensor var_594 = const()[name = string("op_594"), val = tensor([1, 20, 64, -1])]; + tensor var_595_cast_fp16 = reshape(shape = var_594, x = key_7_cast_fp16)[name = string("op_595_cast_fp16")]; + bool mh_w_7_transpose_x_0 = const()[name = string("mh_w_7_transpose_x_0"), val = bool(true)]; + bool mh_w_7_transpose_y_0 = const()[name = string("mh_w_7_transpose_y_0"), val = bool(false)]; + tensor mh_w_7_cast_fp16 = matmul(transpose_x = mh_w_7_transpose_x_0, transpose_y = mh_w_7_transpose_y_0, x = var_593_cast_fp16, y = var_595_cast_fp16)[name = string("mh_w_7_cast_fp16")]; + tensor var_598_cast_fp16 = softmax(axis = var_536, x = mh_w_7_cast_fp16)[name = string("op_598_cast_fp16")]; + tensor var_599 = const()[name = string("op_599"), val = tensor([1, 20, 64, -1])]; + tensor var_600_cast_fp16 = reshape(shape = var_599, x = value_7_cast_fp16)[name = string("op_600_cast_fp16")]; + bool attn_7_transpose_x_0 = const()[name = string("attn_7_transpose_x_0"), val = bool(false)]; + bool attn_7_transpose_y_0 = const()[name = string("attn_7_transpose_y_0"), val = bool(true)]; + tensor attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = var_600_cast_fp16, y = var_598_cast_fp16)[name = string("attn_7_cast_fp16")]; + tensor var_603 = const()[name = string("op_603"), val = tensor([1, 1280, 1, -1])]; + tensor input_25_cast_fp16 = reshape(shape = var_603, x = attn_7_cast_fp16)[name = string("input_25_cast_fp16")]; + string obj_15_pad_type_0 = const()[name = string("obj_15_pad_type_0"), val = string("valid")]; + tensor obj_15_strides_0 = const()[name = string("obj_15_strides_0"), val = tensor([1, 1])]; + tensor obj_15_pad_0 = const()[name = string("obj_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_15_dilations_0 = const()[name = string("obj_15_dilations_0"), val = tensor([1, 1])]; + int32 obj_15_groups_0 = const()[name = string("obj_15_groups_0"), val = int32(1)]; + tensor layers_3_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142565120)))]; + tensor layers_3_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_3_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145841984)))]; + tensor obj_15_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_bias_to_fp16, dilations = obj_15_dilations_0, groups = obj_15_groups_0, pad = obj_15_pad_0, pad_type = obj_15_pad_type_0, strides = obj_15_strides_0, weight = layers_3_self_attn_o_proj_weight_to_fp16, x = input_25_cast_fp16)[name = string("obj_15_cast_fp16")]; + tensor inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = obj_15_cast_fp16)[name = string("inputs_15_cast_fp16")]; + tensor out_15_axes_0 = const()[name = string("out_15_axes_0"), val = tensor([1])]; + fp16 var_621_to_fp16 = const()[name = string("op_621_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_621_to_fp16, x = inputs_15_cast_fp16)[name = string("out_15_cast_fp16")]; + tensor input_27_gamma_0_to_fp16 = const()[name = string("input_27_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145844608)))]; + tensor input_27_beta_0_to_fp16 = const()[name = string("input_27_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145847232)))]; + fp16 input_27_epsilon_0_to_fp16 = const()[name = string("input_27_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor input_27_cast_fp16 = batch_norm(beta = input_27_beta_0_to_fp16, epsilon = input_27_epsilon_0_to_fp16, gamma = input_27_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_15_cast_fp16)[name = string("input_27_cast_fp16")]; + string input_29_pad_type_0 = const()[name = string("input_29_pad_type_0"), val = string("valid")]; + tensor input_29_strides_0 = const()[name = string("input_29_strides_0"), val = tensor([1, 1])]; + tensor input_29_pad_0 = const()[name = string("input_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_29_dilations_0 = const()[name = string("input_29_dilations_0"), val = tensor([1, 1])]; + int32 input_29_groups_0 = const()[name = string("input_29_groups_0"), val = int32(1)]; + tensor layers_3_fc1_weight_to_fp16 = const()[name = string("layers_3_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145849856)))]; + tensor layers_3_fc1_bias_to_fp16 = const()[name = string("layers_3_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158957120)))]; + tensor input_29_cast_fp16 = conv(bias = layers_3_fc1_bias_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 = layers_3_fc1_weight_to_fp16, x = input_27_cast_fp16)[name = string("input_29_cast_fp16")]; + string input_31_mode_0 = const()[name = string("input_31_mode_0"), val = string("EXACT")]; + tensor input_31_cast_fp16 = gelu(mode = input_31_mode_0, x = input_29_cast_fp16)[name = string("input_31_cast_fp16")]; + string hidden_states_11_pad_type_0 = const()[name = string("hidden_states_11_pad_type_0"), val = string("valid")]; + tensor hidden_states_11_strides_0 = const()[name = string("hidden_states_11_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_11_pad_0 = const()[name = string("hidden_states_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_11_dilations_0 = const()[name = string("hidden_states_11_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_11_groups_0 = const()[name = string("hidden_states_11_groups_0"), val = int32(1)]; + tensor layers_3_fc2_weight_to_fp16 = const()[name = string("layers_3_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158967424)))]; + tensor layers_3_fc2_bias_to_fp16 = const()[name = string("layers_3_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172074688)))]; + tensor hidden_states_11_cast_fp16 = conv(bias = layers_3_fc2_bias_to_fp16, dilations = hidden_states_11_dilations_0, groups = hidden_states_11_groups_0, pad = hidden_states_11_pad_0, pad_type = hidden_states_11_pad_type_0, strides = hidden_states_11_strides_0, weight = layers_3_fc2_weight_to_fp16, x = input_31_cast_fp16)[name = string("hidden_states_11_cast_fp16")]; + tensor inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = hidden_states_11_cast_fp16)[name = string("inputs_17_cast_fp16")]; + int32 var_654 = const()[name = string("op_654"), val = int32(3)]; + tensor out_17_axes_0 = const()[name = string("out_17_axes_0"), val = tensor([1])]; + fp16 var_673_to_fp16 = const()[name = string("op_673_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_673_to_fp16, x = inputs_17_cast_fp16)[name = string("out_17_cast_fp16")]; + tensor obj_17_gamma_0_to_fp16 = const()[name = string("obj_17_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172077312)))]; + tensor obj_17_beta_0_to_fp16 = const()[name = string("obj_17_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172079936)))]; + fp16 obj_17_epsilon_0_to_fp16 = const()[name = string("obj_17_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor obj_17_cast_fp16 = batch_norm(beta = obj_17_beta_0_to_fp16, epsilon = obj_17_epsilon_0_to_fp16, gamma = obj_17_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_17_cast_fp16)[name = string("obj_17_cast_fp16")]; + string query_9_pad_type_0 = const()[name = string("query_9_pad_type_0"), val = string("valid")]; + tensor query_9_strides_0 = const()[name = string("query_9_strides_0"), val = tensor([1, 1])]; + tensor query_9_pad_0 = const()[name = string("query_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_9_dilations_0 = const()[name = string("query_9_dilations_0"), val = tensor([1, 1])]; + int32 query_9_groups_0 = const()[name = string("query_9_groups_0"), val = int32(1)]; + tensor layers_4_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172082560)))]; + tensor layers_4_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_4_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175359424)))]; + tensor query_9_cast_fp16 = conv(bias = layers_4_self_attn_q_proj_bias_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_4_self_attn_q_proj_weight_to_fp16, x = obj_17_cast_fp16)[name = string("query_9_cast_fp16")]; + string key_9_pad_type_0 = const()[name = string("key_9_pad_type_0"), val = string("valid")]; + tensor key_9_strides_0 = const()[name = string("key_9_strides_0"), val = tensor([1, 1])]; + tensor key_9_pad_0 = const()[name = string("key_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_9_dilations_0 = const()[name = string("key_9_dilations_0"), val = tensor([1, 1])]; + int32 key_9_groups_0 = const()[name = string("key_9_groups_0"), val = int32(1)]; + tensor layers_4_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175362048)))]; + tensor key_9_cast_fp16 = conv(dilations = key_9_dilations_0, groups = key_9_groups_0, pad = key_9_pad_0, pad_type = key_9_pad_type_0, strides = key_9_strides_0, weight = layers_4_self_attn_k_proj_weight_to_fp16, x = obj_17_cast_fp16)[name = string("key_9_cast_fp16")]; + string value_9_pad_type_0 = const()[name = string("value_9_pad_type_0"), val = string("valid")]; + tensor value_9_strides_0 = const()[name = string("value_9_strides_0"), val = tensor([1, 1])]; + tensor value_9_pad_0 = const()[name = string("value_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_9_dilations_0 = const()[name = string("value_9_dilations_0"), val = tensor([1, 1])]; + int32 value_9_groups_0 = const()[name = string("value_9_groups_0"), val = int32(1)]; + tensor layers_4_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178638912)))]; + tensor layers_4_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_4_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181915776)))]; + tensor value_9_cast_fp16 = conv(bias = layers_4_self_attn_v_proj_bias_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_17_cast_fp16)[name = string("value_9_cast_fp16")]; + tensor var_708 = const()[name = string("op_708"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_9_cast_fp16 = reshape(shape = var_708, x = query_9_cast_fp16)[name = string("mh_q_9_cast_fp16")]; + fp16 var_710_to_fp16 = const()[name = string("op_710_to_fp16"), val = fp16(0x1p-3)]; + tensor var_711_cast_fp16 = mul(x = mh_q_9_cast_fp16, y = var_710_to_fp16)[name = string("op_711_cast_fp16")]; + tensor var_712 = const()[name = string("op_712"), val = tensor([1, 20, 64, -1])]; + tensor var_713_cast_fp16 = reshape(shape = var_712, x = key_9_cast_fp16)[name = string("op_713_cast_fp16")]; + bool mh_w_9_transpose_x_0 = const()[name = string("mh_w_9_transpose_x_0"), val = bool(true)]; + bool mh_w_9_transpose_y_0 = const()[name = string("mh_w_9_transpose_y_0"), val = bool(false)]; + tensor mh_w_9_cast_fp16 = matmul(transpose_x = mh_w_9_transpose_x_0, transpose_y = mh_w_9_transpose_y_0, x = var_711_cast_fp16, y = var_713_cast_fp16)[name = string("mh_w_9_cast_fp16")]; + tensor var_716_cast_fp16 = softmax(axis = var_654, x = mh_w_9_cast_fp16)[name = string("op_716_cast_fp16")]; + tensor var_717 = const()[name = string("op_717"), val = tensor([1, 20, 64, -1])]; + tensor var_718_cast_fp16 = reshape(shape = var_717, x = value_9_cast_fp16)[name = string("op_718_cast_fp16")]; + bool attn_9_transpose_x_0 = const()[name = string("attn_9_transpose_x_0"), val = bool(false)]; + bool attn_9_transpose_y_0 = const()[name = string("attn_9_transpose_y_0"), val = bool(true)]; + tensor attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = var_718_cast_fp16, y = var_716_cast_fp16)[name = string("attn_9_cast_fp16")]; + tensor var_721 = const()[name = string("op_721"), val = tensor([1, 1280, 1, -1])]; + tensor input_33_cast_fp16 = reshape(shape = var_721, x = attn_9_cast_fp16)[name = string("input_33_cast_fp16")]; + string obj_19_pad_type_0 = const()[name = string("obj_19_pad_type_0"), val = string("valid")]; + tensor obj_19_strides_0 = const()[name = string("obj_19_strides_0"), val = tensor([1, 1])]; + tensor obj_19_pad_0 = const()[name = string("obj_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_19_dilations_0 = const()[name = string("obj_19_dilations_0"), val = tensor([1, 1])]; + int32 obj_19_groups_0 = const()[name = string("obj_19_groups_0"), val = int32(1)]; + tensor layers_4_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181918400)))]; + tensor layers_4_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_4_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185195264)))]; + tensor obj_19_cast_fp16 = conv(bias = layers_4_self_attn_o_proj_bias_to_fp16, dilations = obj_19_dilations_0, groups = obj_19_groups_0, pad = obj_19_pad_0, pad_type = obj_19_pad_type_0, strides = obj_19_strides_0, weight = layers_4_self_attn_o_proj_weight_to_fp16, x = input_33_cast_fp16)[name = string("obj_19_cast_fp16")]; + tensor inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = obj_19_cast_fp16)[name = string("inputs_19_cast_fp16")]; + tensor out_19_axes_0 = const()[name = string("out_19_axes_0"), val = tensor([1])]; + fp16 var_739_to_fp16 = const()[name = string("op_739_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_739_to_fp16, x = inputs_19_cast_fp16)[name = string("out_19_cast_fp16")]; + tensor input_35_gamma_0_to_fp16 = const()[name = string("input_35_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185197888)))]; + tensor input_35_beta_0_to_fp16 = const()[name = string("input_35_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185200512)))]; + fp16 input_35_epsilon_0_to_fp16 = const()[name = string("input_35_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor 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 = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_19_cast_fp16)[name = string("input_35_cast_fp16")]; + string input_37_pad_type_0 = const()[name = string("input_37_pad_type_0"), val = string("valid")]; + tensor input_37_strides_0 = const()[name = string("input_37_strides_0"), val = tensor([1, 1])]; + tensor input_37_pad_0 = const()[name = string("input_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_37_dilations_0 = const()[name = string("input_37_dilations_0"), val = tensor([1, 1])]; + int32 input_37_groups_0 = const()[name = string("input_37_groups_0"), val = int32(1)]; + tensor layers_4_fc1_weight_to_fp16 = const()[name = string("layers_4_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185203136)))]; + tensor layers_4_fc1_bias_to_fp16 = const()[name = string("layers_4_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198310400)))]; + tensor input_37_cast_fp16 = conv(bias = layers_4_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_4_fc1_weight_to_fp16, x = input_35_cast_fp16)[name = string("input_37_cast_fp16")]; + string input_39_mode_0 = const()[name = string("input_39_mode_0"), val = string("EXACT")]; + tensor input_39_cast_fp16 = gelu(mode = input_39_mode_0, x = input_37_cast_fp16)[name = string("input_39_cast_fp16")]; + string hidden_states_13_pad_type_0 = const()[name = string("hidden_states_13_pad_type_0"), val = string("valid")]; + tensor hidden_states_13_strides_0 = const()[name = string("hidden_states_13_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_13_pad_0 = const()[name = string("hidden_states_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_13_dilations_0 = const()[name = string("hidden_states_13_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_13_groups_0 = const()[name = string("hidden_states_13_groups_0"), val = int32(1)]; + tensor layers_4_fc2_weight_to_fp16 = const()[name = string("layers_4_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198320704)))]; + tensor layers_4_fc2_bias_to_fp16 = const()[name = string("layers_4_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211427968)))]; + tensor hidden_states_13_cast_fp16 = conv(bias = layers_4_fc2_bias_to_fp16, dilations = hidden_states_13_dilations_0, groups = hidden_states_13_groups_0, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = hidden_states_13_strides_0, weight = layers_4_fc2_weight_to_fp16, x = input_39_cast_fp16)[name = string("hidden_states_13_cast_fp16")]; + tensor inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = hidden_states_13_cast_fp16)[name = string("inputs_21_cast_fp16")]; + int32 var_772 = const()[name = string("op_772"), val = int32(3)]; + tensor out_21_axes_0 = const()[name = string("out_21_axes_0"), val = tensor([1])]; + fp16 var_791_to_fp16 = const()[name = string("op_791_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_791_to_fp16, x = inputs_21_cast_fp16)[name = string("out_21_cast_fp16")]; + tensor obj_21_gamma_0_to_fp16 = const()[name = string("obj_21_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211430592)))]; + tensor obj_21_beta_0_to_fp16 = const()[name = string("obj_21_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211433216)))]; + fp16 obj_21_epsilon_0_to_fp16 = const()[name = string("obj_21_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor obj_21_cast_fp16 = batch_norm(beta = obj_21_beta_0_to_fp16, epsilon = obj_21_epsilon_0_to_fp16, gamma = obj_21_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_21_cast_fp16)[name = string("obj_21_cast_fp16")]; + string query_11_pad_type_0 = const()[name = string("query_11_pad_type_0"), val = string("valid")]; + tensor query_11_strides_0 = const()[name = string("query_11_strides_0"), val = tensor([1, 1])]; + tensor query_11_pad_0 = const()[name = string("query_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_11_dilations_0 = const()[name = string("query_11_dilations_0"), val = tensor([1, 1])]; + int32 query_11_groups_0 = const()[name = string("query_11_groups_0"), val = int32(1)]; + tensor layers_5_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211435840)))]; + tensor layers_5_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_5_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214712704)))]; + tensor query_11_cast_fp16 = conv(bias = layers_5_self_attn_q_proj_bias_to_fp16, dilations = query_11_dilations_0, groups = query_11_groups_0, pad = query_11_pad_0, pad_type = query_11_pad_type_0, strides = query_11_strides_0, weight = layers_5_self_attn_q_proj_weight_to_fp16, x = obj_21_cast_fp16)[name = string("query_11_cast_fp16")]; + string key_11_pad_type_0 = const()[name = string("key_11_pad_type_0"), val = string("valid")]; + tensor key_11_strides_0 = const()[name = string("key_11_strides_0"), val = tensor([1, 1])]; + tensor key_11_pad_0 = const()[name = string("key_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_11_dilations_0 = const()[name = string("key_11_dilations_0"), val = tensor([1, 1])]; + int32 key_11_groups_0 = const()[name = string("key_11_groups_0"), val = int32(1)]; + tensor layers_5_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214715328)))]; + tensor key_11_cast_fp16 = conv(dilations = key_11_dilations_0, groups = key_11_groups_0, pad = key_11_pad_0, pad_type = key_11_pad_type_0, strides = key_11_strides_0, weight = layers_5_self_attn_k_proj_weight_to_fp16, x = obj_21_cast_fp16)[name = string("key_11_cast_fp16")]; + string value_11_pad_type_0 = const()[name = string("value_11_pad_type_0"), val = string("valid")]; + tensor value_11_strides_0 = const()[name = string("value_11_strides_0"), val = tensor([1, 1])]; + tensor value_11_pad_0 = const()[name = string("value_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_11_dilations_0 = const()[name = string("value_11_dilations_0"), val = tensor([1, 1])]; + int32 value_11_groups_0 = const()[name = string("value_11_groups_0"), val = int32(1)]; + tensor layers_5_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217992192)))]; + tensor layers_5_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_5_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221269056)))]; + tensor value_11_cast_fp16 = conv(bias = layers_5_self_attn_v_proj_bias_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_21_cast_fp16)[name = string("value_11_cast_fp16")]; + tensor var_826 = const()[name = string("op_826"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_11_cast_fp16 = reshape(shape = var_826, x = query_11_cast_fp16)[name = string("mh_q_11_cast_fp16")]; + fp16 var_828_to_fp16 = const()[name = string("op_828_to_fp16"), val = fp16(0x1p-3)]; + tensor var_829_cast_fp16 = mul(x = mh_q_11_cast_fp16, y = var_828_to_fp16)[name = string("op_829_cast_fp16")]; + tensor var_830 = const()[name = string("op_830"), val = tensor([1, 20, 64, -1])]; + tensor var_831_cast_fp16 = reshape(shape = var_830, x = key_11_cast_fp16)[name = string("op_831_cast_fp16")]; + bool mh_w_11_transpose_x_0 = const()[name = string("mh_w_11_transpose_x_0"), val = bool(true)]; + bool mh_w_11_transpose_y_0 = const()[name = string("mh_w_11_transpose_y_0"), val = bool(false)]; + tensor mh_w_11_cast_fp16 = matmul(transpose_x = mh_w_11_transpose_x_0, transpose_y = mh_w_11_transpose_y_0, x = var_829_cast_fp16, y = var_831_cast_fp16)[name = string("mh_w_11_cast_fp16")]; + tensor var_834_cast_fp16 = softmax(axis = var_772, x = mh_w_11_cast_fp16)[name = string("op_834_cast_fp16")]; + tensor var_835 = const()[name = string("op_835"), val = tensor([1, 20, 64, -1])]; + tensor var_836_cast_fp16 = reshape(shape = var_835, x = value_11_cast_fp16)[name = string("op_836_cast_fp16")]; + bool attn_11_transpose_x_0 = const()[name = string("attn_11_transpose_x_0"), val = bool(false)]; + bool attn_11_transpose_y_0 = const()[name = string("attn_11_transpose_y_0"), val = bool(true)]; + tensor attn_11_cast_fp16 = matmul(transpose_x = attn_11_transpose_x_0, transpose_y = attn_11_transpose_y_0, x = var_836_cast_fp16, y = var_834_cast_fp16)[name = string("attn_11_cast_fp16")]; + tensor var_839 = const()[name = string("op_839"), val = tensor([1, 1280, 1, -1])]; + tensor input_41_cast_fp16 = reshape(shape = var_839, x = attn_11_cast_fp16)[name = string("input_41_cast_fp16")]; + string obj_23_pad_type_0 = const()[name = string("obj_23_pad_type_0"), val = string("valid")]; + tensor obj_23_strides_0 = const()[name = string("obj_23_strides_0"), val = tensor([1, 1])]; + tensor obj_23_pad_0 = const()[name = string("obj_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_23_dilations_0 = const()[name = string("obj_23_dilations_0"), val = tensor([1, 1])]; + int32 obj_23_groups_0 = const()[name = string("obj_23_groups_0"), val = int32(1)]; + tensor layers_5_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221271680)))]; + tensor layers_5_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_5_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224548544)))]; + tensor obj_23_cast_fp16 = conv(bias = layers_5_self_attn_o_proj_bias_to_fp16, dilations = obj_23_dilations_0, groups = obj_23_groups_0, pad = obj_23_pad_0, pad_type = obj_23_pad_type_0, strides = obj_23_strides_0, weight = layers_5_self_attn_o_proj_weight_to_fp16, x = input_41_cast_fp16)[name = string("obj_23_cast_fp16")]; + tensor inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = obj_23_cast_fp16)[name = string("inputs_23_cast_fp16")]; + tensor out_23_axes_0 = const()[name = string("out_23_axes_0"), val = tensor([1])]; + fp16 var_857_to_fp16 = const()[name = string("op_857_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_857_to_fp16, x = inputs_23_cast_fp16)[name = string("out_23_cast_fp16")]; + tensor input_43_gamma_0_to_fp16 = const()[name = string("input_43_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224551168)))]; + tensor input_43_beta_0_to_fp16 = const()[name = string("input_43_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224553792)))]; + fp16 input_43_epsilon_0_to_fp16 = const()[name = string("input_43_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor 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 = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_23_cast_fp16)[name = string("input_43_cast_fp16")]; + string input_45_pad_type_0 = const()[name = string("input_45_pad_type_0"), val = string("valid")]; + tensor input_45_strides_0 = const()[name = string("input_45_strides_0"), val = tensor([1, 1])]; + tensor input_45_pad_0 = const()[name = string("input_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_45_dilations_0 = const()[name = string("input_45_dilations_0"), val = tensor([1, 1])]; + int32 input_45_groups_0 = const()[name = string("input_45_groups_0"), val = int32(1)]; + tensor layers_5_fc1_weight_to_fp16 = const()[name = string("layers_5_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224556416)))]; + tensor layers_5_fc1_bias_to_fp16 = const()[name = string("layers_5_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237663680)))]; + tensor input_45_cast_fp16 = conv(bias = layers_5_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_5_fc1_weight_to_fp16, x = input_43_cast_fp16)[name = string("input_45_cast_fp16")]; + string input_47_mode_0 = const()[name = string("input_47_mode_0"), val = string("EXACT")]; + tensor input_47_cast_fp16 = gelu(mode = input_47_mode_0, x = input_45_cast_fp16)[name = string("input_47_cast_fp16")]; + string hidden_states_15_pad_type_0 = const()[name = string("hidden_states_15_pad_type_0"), val = string("valid")]; + tensor hidden_states_15_strides_0 = const()[name = string("hidden_states_15_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_15_pad_0 = const()[name = string("hidden_states_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_15_dilations_0 = const()[name = string("hidden_states_15_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_15_groups_0 = const()[name = string("hidden_states_15_groups_0"), val = int32(1)]; + tensor layers_5_fc2_weight_to_fp16 = const()[name = string("layers_5_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237673984)))]; + tensor layers_5_fc2_bias_to_fp16 = const()[name = string("layers_5_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250781248)))]; + tensor hidden_states_15_cast_fp16 = conv(bias = layers_5_fc2_bias_to_fp16, dilations = hidden_states_15_dilations_0, groups = hidden_states_15_groups_0, pad = hidden_states_15_pad_0, pad_type = hidden_states_15_pad_type_0, strides = hidden_states_15_strides_0, weight = layers_5_fc2_weight_to_fp16, x = input_47_cast_fp16)[name = string("hidden_states_15_cast_fp16")]; + tensor inputs_25_cast_fp16 = add(x = inputs_23_cast_fp16, y = hidden_states_15_cast_fp16)[name = string("inputs_25_cast_fp16")]; + int32 var_890 = const()[name = string("op_890"), val = int32(3)]; + tensor out_25_axes_0 = const()[name = string("out_25_axes_0"), val = tensor([1])]; + fp16 var_909_to_fp16 = const()[name = string("op_909_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_25_cast_fp16 = layer_norm(axes = out_25_axes_0, epsilon = var_909_to_fp16, x = inputs_25_cast_fp16)[name = string("out_25_cast_fp16")]; + tensor obj_25_gamma_0_to_fp16 = const()[name = string("obj_25_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250783872)))]; + tensor obj_25_beta_0_to_fp16 = const()[name = string("obj_25_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250786496)))]; + fp16 obj_25_epsilon_0_to_fp16 = const()[name = string("obj_25_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor obj_25_cast_fp16 = batch_norm(beta = obj_25_beta_0_to_fp16, epsilon = obj_25_epsilon_0_to_fp16, gamma = obj_25_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_25_cast_fp16)[name = string("obj_25_cast_fp16")]; + string query_13_pad_type_0 = const()[name = string("query_13_pad_type_0"), val = string("valid")]; + tensor query_13_strides_0 = const()[name = string("query_13_strides_0"), val = tensor([1, 1])]; + tensor query_13_pad_0 = const()[name = string("query_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_13_dilations_0 = const()[name = string("query_13_dilations_0"), val = tensor([1, 1])]; + int32 query_13_groups_0 = const()[name = string("query_13_groups_0"), val = int32(1)]; + tensor layers_6_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250789120)))]; + tensor layers_6_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_6_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254065984)))]; + tensor query_13_cast_fp16 = conv(bias = layers_6_self_attn_q_proj_bias_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_6_self_attn_q_proj_weight_to_fp16, x = obj_25_cast_fp16)[name = string("query_13_cast_fp16")]; + string key_13_pad_type_0 = const()[name = string("key_13_pad_type_0"), val = string("valid")]; + tensor key_13_strides_0 = const()[name = string("key_13_strides_0"), val = tensor([1, 1])]; + tensor key_13_pad_0 = const()[name = string("key_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_13_dilations_0 = const()[name = string("key_13_dilations_0"), val = tensor([1, 1])]; + int32 key_13_groups_0 = const()[name = string("key_13_groups_0"), val = int32(1)]; + tensor layers_6_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254068608)))]; + tensor key_13_cast_fp16 = conv(dilations = key_13_dilations_0, groups = key_13_groups_0, pad = key_13_pad_0, pad_type = key_13_pad_type_0, strides = key_13_strides_0, weight = layers_6_self_attn_k_proj_weight_to_fp16, x = obj_25_cast_fp16)[name = string("key_13_cast_fp16")]; + string value_13_pad_type_0 = const()[name = string("value_13_pad_type_0"), val = string("valid")]; + tensor value_13_strides_0 = const()[name = string("value_13_strides_0"), val = tensor([1, 1])]; + tensor value_13_pad_0 = const()[name = string("value_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_13_dilations_0 = const()[name = string("value_13_dilations_0"), val = tensor([1, 1])]; + int32 value_13_groups_0 = const()[name = string("value_13_groups_0"), val = int32(1)]; + tensor layers_6_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257345472)))]; + tensor layers_6_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_6_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260622336)))]; + tensor value_13_cast_fp16 = conv(bias = layers_6_self_attn_v_proj_bias_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_25_cast_fp16)[name = string("value_13_cast_fp16")]; + tensor var_944 = const()[name = string("op_944"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_13_cast_fp16 = reshape(shape = var_944, x = query_13_cast_fp16)[name = string("mh_q_13_cast_fp16")]; + fp16 var_946_to_fp16 = const()[name = string("op_946_to_fp16"), val = fp16(0x1p-3)]; + tensor var_947_cast_fp16 = mul(x = mh_q_13_cast_fp16, y = var_946_to_fp16)[name = string("op_947_cast_fp16")]; + tensor var_948 = const()[name = string("op_948"), val = tensor([1, 20, 64, -1])]; + tensor var_949_cast_fp16 = reshape(shape = var_948, x = key_13_cast_fp16)[name = string("op_949_cast_fp16")]; + bool mh_w_13_transpose_x_0 = const()[name = string("mh_w_13_transpose_x_0"), val = bool(true)]; + bool mh_w_13_transpose_y_0 = const()[name = string("mh_w_13_transpose_y_0"), val = bool(false)]; + tensor mh_w_13_cast_fp16 = matmul(transpose_x = mh_w_13_transpose_x_0, transpose_y = mh_w_13_transpose_y_0, x = var_947_cast_fp16, y = var_949_cast_fp16)[name = string("mh_w_13_cast_fp16")]; + tensor var_952_cast_fp16 = softmax(axis = var_890, x = mh_w_13_cast_fp16)[name = string("op_952_cast_fp16")]; + tensor var_953 = const()[name = string("op_953"), val = tensor([1, 20, 64, -1])]; + tensor var_954_cast_fp16 = reshape(shape = var_953, x = value_13_cast_fp16)[name = string("op_954_cast_fp16")]; + bool attn_13_transpose_x_0 = const()[name = string("attn_13_transpose_x_0"), val = bool(false)]; + bool attn_13_transpose_y_0 = const()[name = string("attn_13_transpose_y_0"), val = bool(true)]; + tensor attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = var_954_cast_fp16, y = var_952_cast_fp16)[name = string("attn_13_cast_fp16")]; + tensor var_957 = const()[name = string("op_957"), val = tensor([1, 1280, 1, -1])]; + tensor input_49_cast_fp16 = reshape(shape = var_957, x = attn_13_cast_fp16)[name = string("input_49_cast_fp16")]; + string obj_27_pad_type_0 = const()[name = string("obj_27_pad_type_0"), val = string("valid")]; + tensor obj_27_strides_0 = const()[name = string("obj_27_strides_0"), val = tensor([1, 1])]; + tensor obj_27_pad_0 = const()[name = string("obj_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_27_dilations_0 = const()[name = string("obj_27_dilations_0"), val = tensor([1, 1])]; + int32 obj_27_groups_0 = const()[name = string("obj_27_groups_0"), val = int32(1)]; + tensor layers_6_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260624960)))]; + tensor layers_6_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_6_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263901824)))]; + tensor obj_27_cast_fp16 = conv(bias = layers_6_self_attn_o_proj_bias_to_fp16, dilations = obj_27_dilations_0, groups = obj_27_groups_0, pad = obj_27_pad_0, pad_type = obj_27_pad_type_0, strides = obj_27_strides_0, weight = layers_6_self_attn_o_proj_weight_to_fp16, x = input_49_cast_fp16)[name = string("obj_27_cast_fp16")]; + tensor inputs_27_cast_fp16 = add(x = inputs_25_cast_fp16, y = obj_27_cast_fp16)[name = string("inputs_27_cast_fp16")]; + tensor out_27_axes_0 = const()[name = string("out_27_axes_0"), val = tensor([1])]; + fp16 var_975_to_fp16 = const()[name = string("op_975_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_27_cast_fp16 = layer_norm(axes = out_27_axes_0, epsilon = var_975_to_fp16, x = inputs_27_cast_fp16)[name = string("out_27_cast_fp16")]; + tensor input_51_gamma_0_to_fp16 = const()[name = string("input_51_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263904448)))]; + tensor input_51_beta_0_to_fp16 = const()[name = string("input_51_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263907072)))]; + fp16 input_51_epsilon_0_to_fp16 = const()[name = string("input_51_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor 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 = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_27_cast_fp16)[name = string("input_51_cast_fp16")]; + string input_53_pad_type_0 = const()[name = string("input_53_pad_type_0"), val = string("valid")]; + tensor input_53_strides_0 = const()[name = string("input_53_strides_0"), val = tensor([1, 1])]; + tensor input_53_pad_0 = const()[name = string("input_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_53_dilations_0 = const()[name = string("input_53_dilations_0"), val = tensor([1, 1])]; + int32 input_53_groups_0 = const()[name = string("input_53_groups_0"), val = int32(1)]; + tensor layers_6_fc1_weight_to_fp16 = const()[name = string("layers_6_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263909696)))]; + tensor layers_6_fc1_bias_to_fp16 = const()[name = string("layers_6_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277016960)))]; + tensor input_53_cast_fp16 = conv(bias = layers_6_fc1_bias_to_fp16, 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_6_fc1_weight_to_fp16, x = input_51_cast_fp16)[name = string("input_53_cast_fp16")]; + string input_55_mode_0 = const()[name = string("input_55_mode_0"), val = string("EXACT")]; + tensor input_55_cast_fp16 = gelu(mode = input_55_mode_0, x = input_53_cast_fp16)[name = string("input_55_cast_fp16")]; + string hidden_states_17_pad_type_0 = const()[name = string("hidden_states_17_pad_type_0"), val = string("valid")]; + tensor hidden_states_17_strides_0 = const()[name = string("hidden_states_17_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_17_pad_0 = const()[name = string("hidden_states_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_17_dilations_0 = const()[name = string("hidden_states_17_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_17_groups_0 = const()[name = string("hidden_states_17_groups_0"), val = int32(1)]; + tensor layers_6_fc2_weight_to_fp16 = const()[name = string("layers_6_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277027264)))]; + tensor layers_6_fc2_bias_to_fp16 = const()[name = string("layers_6_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290134528)))]; + tensor hidden_states_17_cast_fp16 = conv(bias = layers_6_fc2_bias_to_fp16, dilations = hidden_states_17_dilations_0, groups = hidden_states_17_groups_0, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = hidden_states_17_strides_0, weight = layers_6_fc2_weight_to_fp16, x = input_55_cast_fp16)[name = string("hidden_states_17_cast_fp16")]; + tensor inputs_29_cast_fp16 = add(x = inputs_27_cast_fp16, y = hidden_states_17_cast_fp16)[name = string("inputs_29_cast_fp16")]; + int32 var_1008 = const()[name = string("op_1008"), val = int32(3)]; + tensor out_29_axes_0 = const()[name = string("out_29_axes_0"), val = tensor([1])]; + fp16 var_1027_to_fp16 = const()[name = string("op_1027_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_29_cast_fp16 = layer_norm(axes = out_29_axes_0, epsilon = var_1027_to_fp16, x = inputs_29_cast_fp16)[name = string("out_29_cast_fp16")]; + tensor obj_29_gamma_0_to_fp16 = const()[name = string("obj_29_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290137152)))]; + tensor obj_29_beta_0_to_fp16 = const()[name = string("obj_29_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290139776)))]; + fp16 obj_29_epsilon_0_to_fp16 = const()[name = string("obj_29_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor obj_29_cast_fp16 = batch_norm(beta = obj_29_beta_0_to_fp16, epsilon = obj_29_epsilon_0_to_fp16, gamma = obj_29_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_29_cast_fp16)[name = string("obj_29_cast_fp16")]; + string query_15_pad_type_0 = const()[name = string("query_15_pad_type_0"), val = string("valid")]; + tensor query_15_strides_0 = const()[name = string("query_15_strides_0"), val = tensor([1, 1])]; + tensor query_15_pad_0 = const()[name = string("query_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_15_dilations_0 = const()[name = string("query_15_dilations_0"), val = tensor([1, 1])]; + int32 query_15_groups_0 = const()[name = string("query_15_groups_0"), val = int32(1)]; + tensor layers_7_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290142400)))]; + tensor layers_7_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_7_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293419264)))]; + tensor query_15_cast_fp16 = conv(bias = layers_7_self_attn_q_proj_bias_to_fp16, dilations = query_15_dilations_0, groups = query_15_groups_0, pad = query_15_pad_0, pad_type = query_15_pad_type_0, strides = query_15_strides_0, weight = layers_7_self_attn_q_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = string("query_15_cast_fp16")]; + string key_15_pad_type_0 = const()[name = string("key_15_pad_type_0"), val = string("valid")]; + tensor key_15_strides_0 = const()[name = string("key_15_strides_0"), val = tensor([1, 1])]; + tensor key_15_pad_0 = const()[name = string("key_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_15_dilations_0 = const()[name = string("key_15_dilations_0"), val = tensor([1, 1])]; + int32 key_15_groups_0 = const()[name = string("key_15_groups_0"), val = int32(1)]; + tensor layers_7_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293421888)))]; + tensor key_15_cast_fp16 = conv(dilations = key_15_dilations_0, groups = key_15_groups_0, pad = key_15_pad_0, pad_type = key_15_pad_type_0, strides = key_15_strides_0, weight = layers_7_self_attn_k_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = string("key_15_cast_fp16")]; + string value_15_pad_type_0 = const()[name = string("value_15_pad_type_0"), val = string("valid")]; + tensor value_15_strides_0 = const()[name = string("value_15_strides_0"), val = tensor([1, 1])]; + tensor value_15_pad_0 = const()[name = string("value_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_15_dilations_0 = const()[name = string("value_15_dilations_0"), val = tensor([1, 1])]; + int32 value_15_groups_0 = const()[name = string("value_15_groups_0"), val = int32(1)]; + tensor layers_7_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296698752)))]; + tensor layers_7_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_7_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299975616)))]; + tensor value_15_cast_fp16 = conv(bias = layers_7_self_attn_v_proj_bias_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_29_cast_fp16)[name = string("value_15_cast_fp16")]; + tensor var_1062 = const()[name = string("op_1062"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_15_cast_fp16 = reshape(shape = var_1062, x = query_15_cast_fp16)[name = string("mh_q_15_cast_fp16")]; + fp16 var_1064_to_fp16 = const()[name = string("op_1064_to_fp16"), val = fp16(0x1p-3)]; + tensor var_1065_cast_fp16 = mul(x = mh_q_15_cast_fp16, y = var_1064_to_fp16)[name = string("op_1065_cast_fp16")]; + tensor var_1066 = const()[name = string("op_1066"), val = tensor([1, 20, 64, -1])]; + tensor var_1067_cast_fp16 = reshape(shape = var_1066, x = key_15_cast_fp16)[name = string("op_1067_cast_fp16")]; + bool mh_w_15_transpose_x_0 = const()[name = string("mh_w_15_transpose_x_0"), val = bool(true)]; + bool mh_w_15_transpose_y_0 = const()[name = string("mh_w_15_transpose_y_0"), val = bool(false)]; + tensor mh_w_15_cast_fp16 = matmul(transpose_x = mh_w_15_transpose_x_0, transpose_y = mh_w_15_transpose_y_0, x = var_1065_cast_fp16, y = var_1067_cast_fp16)[name = string("mh_w_15_cast_fp16")]; + tensor var_1070_cast_fp16 = softmax(axis = var_1008, x = mh_w_15_cast_fp16)[name = string("op_1070_cast_fp16")]; + tensor var_1071 = const()[name = string("op_1071"), val = tensor([1, 20, 64, -1])]; + tensor var_1072_cast_fp16 = reshape(shape = var_1071, x = value_15_cast_fp16)[name = string("op_1072_cast_fp16")]; + bool attn_15_transpose_x_0 = const()[name = string("attn_15_transpose_x_0"), val = bool(false)]; + bool attn_15_transpose_y_0 = const()[name = string("attn_15_transpose_y_0"), val = bool(true)]; + tensor attn_15_cast_fp16 = matmul(transpose_x = attn_15_transpose_x_0, transpose_y = attn_15_transpose_y_0, x = var_1072_cast_fp16, y = var_1070_cast_fp16)[name = string("attn_15_cast_fp16")]; + tensor var_1075 = const()[name = string("op_1075"), val = tensor([1, 1280, 1, -1])]; + tensor input_57_cast_fp16 = reshape(shape = var_1075, x = attn_15_cast_fp16)[name = string("input_57_cast_fp16")]; + string obj_31_pad_type_0 = const()[name = string("obj_31_pad_type_0"), val = string("valid")]; + tensor obj_31_strides_0 = const()[name = string("obj_31_strides_0"), val = tensor([1, 1])]; + tensor obj_31_pad_0 = const()[name = string("obj_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_31_dilations_0 = const()[name = string("obj_31_dilations_0"), val = tensor([1, 1])]; + int32 obj_31_groups_0 = const()[name = string("obj_31_groups_0"), val = int32(1)]; + tensor layers_7_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299978240)))]; + tensor layers_7_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_7_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303255104)))]; + tensor obj_31_cast_fp16 = conv(bias = layers_7_self_attn_o_proj_bias_to_fp16, dilations = obj_31_dilations_0, groups = obj_31_groups_0, pad = obj_31_pad_0, pad_type = obj_31_pad_type_0, strides = obj_31_strides_0, weight = layers_7_self_attn_o_proj_weight_to_fp16, x = input_57_cast_fp16)[name = string("obj_31_cast_fp16")]; + tensor inputs_31_cast_fp16 = add(x = inputs_29_cast_fp16, y = obj_31_cast_fp16)[name = string("inputs_31_cast_fp16")]; + tensor out_31_axes_0 = const()[name = string("out_31_axes_0"), val = tensor([1])]; + fp16 var_1093_to_fp16 = const()[name = string("op_1093_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_31_cast_fp16 = layer_norm(axes = out_31_axes_0, epsilon = var_1093_to_fp16, x = inputs_31_cast_fp16)[name = string("out_31_cast_fp16")]; + tensor input_59_gamma_0_to_fp16 = const()[name = string("input_59_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303257728)))]; + tensor input_59_beta_0_to_fp16 = const()[name = string("input_59_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303260352)))]; + fp16 input_59_epsilon_0_to_fp16 = const()[name = string("input_59_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor input_59_cast_fp16 = batch_norm(beta = input_59_beta_0_to_fp16, epsilon = input_59_epsilon_0_to_fp16, gamma = input_59_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_31_cast_fp16)[name = string("input_59_cast_fp16")]; + string input_61_pad_type_0 = const()[name = string("input_61_pad_type_0"), val = string("valid")]; + tensor input_61_strides_0 = const()[name = string("input_61_strides_0"), val = tensor([1, 1])]; + tensor input_61_pad_0 = const()[name = string("input_61_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_61_dilations_0 = const()[name = string("input_61_dilations_0"), val = tensor([1, 1])]; + int32 input_61_groups_0 = const()[name = string("input_61_groups_0"), val = int32(1)]; + tensor layers_7_fc1_weight_to_fp16 = const()[name = string("layers_7_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303262976)))]; + tensor layers_7_fc1_bias_to_fp16 = const()[name = string("layers_7_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316370240)))]; + tensor input_61_cast_fp16 = conv(bias = layers_7_fc1_bias_to_fp16, dilations = input_61_dilations_0, groups = input_61_groups_0, pad = input_61_pad_0, pad_type = input_61_pad_type_0, strides = input_61_strides_0, weight = layers_7_fc1_weight_to_fp16, x = input_59_cast_fp16)[name = string("input_61_cast_fp16")]; + string input_63_mode_0 = const()[name = string("input_63_mode_0"), val = string("EXACT")]; + tensor input_63_cast_fp16 = gelu(mode = input_63_mode_0, x = input_61_cast_fp16)[name = string("input_63_cast_fp16")]; + string hidden_states_19_pad_type_0 = const()[name = string("hidden_states_19_pad_type_0"), val = string("valid")]; + tensor hidden_states_19_strides_0 = const()[name = string("hidden_states_19_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_19_pad_0 = const()[name = string("hidden_states_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_19_dilations_0 = const()[name = string("hidden_states_19_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_19_groups_0 = const()[name = string("hidden_states_19_groups_0"), val = int32(1)]; + tensor layers_7_fc2_weight_to_fp16 = const()[name = string("layers_7_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316380544)))]; + tensor layers_7_fc2_bias_to_fp16 = const()[name = string("layers_7_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329487808)))]; + tensor hidden_states_19_cast_fp16 = conv(bias = layers_7_fc2_bias_to_fp16, dilations = hidden_states_19_dilations_0, groups = hidden_states_19_groups_0, pad = hidden_states_19_pad_0, pad_type = hidden_states_19_pad_type_0, strides = hidden_states_19_strides_0, weight = layers_7_fc2_weight_to_fp16, x = input_63_cast_fp16)[name = string("hidden_states_19_cast_fp16")]; + tensor inputs_33_cast_fp16 = add(x = inputs_31_cast_fp16, y = hidden_states_19_cast_fp16)[name = string("inputs_33_cast_fp16")]; + int32 var_1126 = const()[name = string("op_1126"), val = int32(3)]; + tensor out_33_axes_0 = const()[name = string("out_33_axes_0"), val = tensor([1])]; + fp16 var_1145_to_fp16 = const()[name = string("op_1145_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_33_cast_fp16 = layer_norm(axes = out_33_axes_0, epsilon = var_1145_to_fp16, x = inputs_33_cast_fp16)[name = string("out_33_cast_fp16")]; + tensor obj_33_gamma_0_to_fp16 = const()[name = string("obj_33_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329490432)))]; + tensor obj_33_beta_0_to_fp16 = const()[name = string("obj_33_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329493056)))]; + fp16 obj_33_epsilon_0_to_fp16 = const()[name = string("obj_33_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor obj_33_cast_fp16 = batch_norm(beta = obj_33_beta_0_to_fp16, epsilon = obj_33_epsilon_0_to_fp16, gamma = obj_33_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_33_cast_fp16)[name = string("obj_33_cast_fp16")]; + string query_17_pad_type_0 = const()[name = string("query_17_pad_type_0"), val = string("valid")]; + tensor query_17_strides_0 = const()[name = string("query_17_strides_0"), val = tensor([1, 1])]; + tensor query_17_pad_0 = const()[name = string("query_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_17_dilations_0 = const()[name = string("query_17_dilations_0"), val = tensor([1, 1])]; + int32 query_17_groups_0 = const()[name = string("query_17_groups_0"), val = int32(1)]; + tensor layers_8_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329495680)))]; + tensor layers_8_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_8_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332772544)))]; + tensor query_17_cast_fp16 = conv(bias = layers_8_self_attn_q_proj_bias_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_8_self_attn_q_proj_weight_to_fp16, x = obj_33_cast_fp16)[name = string("query_17_cast_fp16")]; + string key_17_pad_type_0 = const()[name = string("key_17_pad_type_0"), val = string("valid")]; + tensor key_17_strides_0 = const()[name = string("key_17_strides_0"), val = tensor([1, 1])]; + tensor key_17_pad_0 = const()[name = string("key_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_17_dilations_0 = const()[name = string("key_17_dilations_0"), val = tensor([1, 1])]; + int32 key_17_groups_0 = const()[name = string("key_17_groups_0"), val = int32(1)]; + tensor layers_8_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332775168)))]; + tensor key_17_cast_fp16 = conv(dilations = key_17_dilations_0, groups = key_17_groups_0, pad = key_17_pad_0, pad_type = key_17_pad_type_0, strides = key_17_strides_0, weight = layers_8_self_attn_k_proj_weight_to_fp16, x = obj_33_cast_fp16)[name = string("key_17_cast_fp16")]; + string value_17_pad_type_0 = const()[name = string("value_17_pad_type_0"), val = string("valid")]; + tensor value_17_strides_0 = const()[name = string("value_17_strides_0"), val = tensor([1, 1])]; + tensor value_17_pad_0 = const()[name = string("value_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_17_dilations_0 = const()[name = string("value_17_dilations_0"), val = tensor([1, 1])]; + int32 value_17_groups_0 = const()[name = string("value_17_groups_0"), val = int32(1)]; + tensor layers_8_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336052032)))]; + tensor layers_8_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_8_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339328896)))]; + tensor value_17_cast_fp16 = conv(bias = layers_8_self_attn_v_proj_bias_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_33_cast_fp16)[name = string("value_17_cast_fp16")]; + tensor var_1180 = const()[name = string("op_1180"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_17_cast_fp16 = reshape(shape = var_1180, x = query_17_cast_fp16)[name = string("mh_q_17_cast_fp16")]; + fp16 var_1182_to_fp16 = const()[name = string("op_1182_to_fp16"), val = fp16(0x1p-3)]; + tensor var_1183_cast_fp16 = mul(x = mh_q_17_cast_fp16, y = var_1182_to_fp16)[name = string("op_1183_cast_fp16")]; + tensor var_1184 = const()[name = string("op_1184"), val = tensor([1, 20, 64, -1])]; + tensor var_1185_cast_fp16 = reshape(shape = var_1184, x = key_17_cast_fp16)[name = string("op_1185_cast_fp16")]; + bool mh_w_17_transpose_x_0 = const()[name = string("mh_w_17_transpose_x_0"), val = bool(true)]; + bool mh_w_17_transpose_y_0 = const()[name = string("mh_w_17_transpose_y_0"), val = bool(false)]; + tensor mh_w_17_cast_fp16 = matmul(transpose_x = mh_w_17_transpose_x_0, transpose_y = mh_w_17_transpose_y_0, x = var_1183_cast_fp16, y = var_1185_cast_fp16)[name = string("mh_w_17_cast_fp16")]; + tensor var_1188_cast_fp16 = softmax(axis = var_1126, x = mh_w_17_cast_fp16)[name = string("op_1188_cast_fp16")]; + tensor var_1189 = const()[name = string("op_1189"), val = tensor([1, 20, 64, -1])]; + tensor var_1190_cast_fp16 = reshape(shape = var_1189, x = value_17_cast_fp16)[name = string("op_1190_cast_fp16")]; + bool attn_17_transpose_x_0 = const()[name = string("attn_17_transpose_x_0"), val = bool(false)]; + bool attn_17_transpose_y_0 = const()[name = string("attn_17_transpose_y_0"), val = bool(true)]; + tensor attn_17_cast_fp16 = matmul(transpose_x = attn_17_transpose_x_0, transpose_y = attn_17_transpose_y_0, x = var_1190_cast_fp16, y = var_1188_cast_fp16)[name = string("attn_17_cast_fp16")]; + tensor var_1193 = const()[name = string("op_1193"), val = tensor([1, 1280, 1, -1])]; + tensor input_65_cast_fp16 = reshape(shape = var_1193, x = attn_17_cast_fp16)[name = string("input_65_cast_fp16")]; + string obj_35_pad_type_0 = const()[name = string("obj_35_pad_type_0"), val = string("valid")]; + tensor obj_35_strides_0 = const()[name = string("obj_35_strides_0"), val = tensor([1, 1])]; + tensor obj_35_pad_0 = const()[name = string("obj_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_35_dilations_0 = const()[name = string("obj_35_dilations_0"), val = tensor([1, 1])]; + int32 obj_35_groups_0 = const()[name = string("obj_35_groups_0"), val = int32(1)]; + tensor layers_8_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339331520)))]; + tensor layers_8_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_8_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342608384)))]; + tensor obj_35_cast_fp16 = conv(bias = layers_8_self_attn_o_proj_bias_to_fp16, dilations = obj_35_dilations_0, groups = obj_35_groups_0, pad = obj_35_pad_0, pad_type = obj_35_pad_type_0, strides = obj_35_strides_0, weight = layers_8_self_attn_o_proj_weight_to_fp16, x = input_65_cast_fp16)[name = string("obj_35_cast_fp16")]; + tensor inputs_35_cast_fp16 = add(x = inputs_33_cast_fp16, y = obj_35_cast_fp16)[name = string("inputs_35_cast_fp16")]; + tensor out_35_axes_0 = const()[name = string("out_35_axes_0"), val = tensor([1])]; + fp16 var_1211_to_fp16 = const()[name = string("op_1211_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_35_cast_fp16 = layer_norm(axes = out_35_axes_0, epsilon = var_1211_to_fp16, x = inputs_35_cast_fp16)[name = string("out_35_cast_fp16")]; + tensor input_67_gamma_0_to_fp16 = const()[name = string("input_67_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342611008)))]; + tensor input_67_beta_0_to_fp16 = const()[name = string("input_67_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342613632)))]; + fp16 input_67_epsilon_0_to_fp16 = const()[name = string("input_67_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor 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 = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_35_cast_fp16)[name = string("input_67_cast_fp16")]; + string input_69_pad_type_0 = const()[name = string("input_69_pad_type_0"), val = string("valid")]; + tensor input_69_strides_0 = const()[name = string("input_69_strides_0"), val = tensor([1, 1])]; + tensor input_69_pad_0 = const()[name = string("input_69_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_69_dilations_0 = const()[name = string("input_69_dilations_0"), val = tensor([1, 1])]; + int32 input_69_groups_0 = const()[name = string("input_69_groups_0"), val = int32(1)]; + tensor layers_8_fc1_weight_to_fp16 = const()[name = string("layers_8_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342616256)))]; + tensor layers_8_fc1_bias_to_fp16 = const()[name = string("layers_8_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355723520)))]; + tensor input_69_cast_fp16 = conv(bias = layers_8_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_8_fc1_weight_to_fp16, x = input_67_cast_fp16)[name = string("input_69_cast_fp16")]; + string input_71_mode_0 = const()[name = string("input_71_mode_0"), val = string("EXACT")]; + tensor input_71_cast_fp16 = gelu(mode = input_71_mode_0, x = input_69_cast_fp16)[name = string("input_71_cast_fp16")]; + string hidden_states_21_pad_type_0 = const()[name = string("hidden_states_21_pad_type_0"), val = string("valid")]; + tensor hidden_states_21_strides_0 = const()[name = string("hidden_states_21_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_21_pad_0 = const()[name = string("hidden_states_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_21_dilations_0 = const()[name = string("hidden_states_21_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_21_groups_0 = const()[name = string("hidden_states_21_groups_0"), val = int32(1)]; + tensor layers_8_fc2_weight_to_fp16 = const()[name = string("layers_8_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355733824)))]; + tensor layers_8_fc2_bias_to_fp16 = const()[name = string("layers_8_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368841088)))]; + tensor hidden_states_21_cast_fp16 = conv(bias = layers_8_fc2_bias_to_fp16, dilations = hidden_states_21_dilations_0, groups = hidden_states_21_groups_0, pad = hidden_states_21_pad_0, pad_type = hidden_states_21_pad_type_0, strides = hidden_states_21_strides_0, weight = layers_8_fc2_weight_to_fp16, x = input_71_cast_fp16)[name = string("hidden_states_21_cast_fp16")]; + tensor inputs_37_cast_fp16 = add(x = inputs_35_cast_fp16, y = hidden_states_21_cast_fp16)[name = string("inputs_37_cast_fp16")]; + int32 var_1244 = const()[name = string("op_1244"), val = int32(3)]; + tensor out_37_axes_0 = const()[name = string("out_37_axes_0"), val = tensor([1])]; + fp16 var_1263_to_fp16 = const()[name = string("op_1263_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_37_cast_fp16 = layer_norm(axes = out_37_axes_0, epsilon = var_1263_to_fp16, x = inputs_37_cast_fp16)[name = string("out_37_cast_fp16")]; + tensor obj_37_gamma_0_to_fp16 = const()[name = string("obj_37_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368843712)))]; + tensor obj_37_beta_0_to_fp16 = const()[name = string("obj_37_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368846336)))]; + fp16 obj_37_epsilon_0_to_fp16 = const()[name = string("obj_37_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor obj_37_cast_fp16 = batch_norm(beta = obj_37_beta_0_to_fp16, epsilon = obj_37_epsilon_0_to_fp16, gamma = obj_37_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_37_cast_fp16)[name = string("obj_37_cast_fp16")]; + string query_19_pad_type_0 = const()[name = string("query_19_pad_type_0"), val = string("valid")]; + tensor query_19_strides_0 = const()[name = string("query_19_strides_0"), val = tensor([1, 1])]; + tensor query_19_pad_0 = const()[name = string("query_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_19_dilations_0 = const()[name = string("query_19_dilations_0"), val = tensor([1, 1])]; + int32 query_19_groups_0 = const()[name = string("query_19_groups_0"), val = int32(1)]; + tensor layers_9_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368848960)))]; + tensor layers_9_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_9_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372125824)))]; + tensor query_19_cast_fp16 = conv(bias = layers_9_self_attn_q_proj_bias_to_fp16, dilations = query_19_dilations_0, groups = query_19_groups_0, pad = query_19_pad_0, pad_type = query_19_pad_type_0, strides = query_19_strides_0, weight = layers_9_self_attn_q_proj_weight_to_fp16, x = obj_37_cast_fp16)[name = string("query_19_cast_fp16")]; + string key_19_pad_type_0 = const()[name = string("key_19_pad_type_0"), val = string("valid")]; + tensor key_19_strides_0 = const()[name = string("key_19_strides_0"), val = tensor([1, 1])]; + tensor key_19_pad_0 = const()[name = string("key_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_19_dilations_0 = const()[name = string("key_19_dilations_0"), val = tensor([1, 1])]; + int32 key_19_groups_0 = const()[name = string("key_19_groups_0"), val = int32(1)]; + tensor layers_9_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372128448)))]; + tensor key_19_cast_fp16 = conv(dilations = key_19_dilations_0, groups = key_19_groups_0, pad = key_19_pad_0, pad_type = key_19_pad_type_0, strides = key_19_strides_0, weight = layers_9_self_attn_k_proj_weight_to_fp16, x = obj_37_cast_fp16)[name = string("key_19_cast_fp16")]; + string value_19_pad_type_0 = const()[name = string("value_19_pad_type_0"), val = string("valid")]; + tensor value_19_strides_0 = const()[name = string("value_19_strides_0"), val = tensor([1, 1])]; + tensor value_19_pad_0 = const()[name = string("value_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_19_dilations_0 = const()[name = string("value_19_dilations_0"), val = tensor([1, 1])]; + int32 value_19_groups_0 = const()[name = string("value_19_groups_0"), val = int32(1)]; + tensor layers_9_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375405312)))]; + tensor layers_9_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_9_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(378682176)))]; + tensor value_19_cast_fp16 = conv(bias = layers_9_self_attn_v_proj_bias_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_37_cast_fp16)[name = string("value_19_cast_fp16")]; + tensor var_1298 = const()[name = string("op_1298"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_19_cast_fp16 = reshape(shape = var_1298, x = query_19_cast_fp16)[name = string("mh_q_19_cast_fp16")]; + fp16 var_1300_to_fp16 = const()[name = string("op_1300_to_fp16"), val = fp16(0x1p-3)]; + tensor var_1301_cast_fp16 = mul(x = mh_q_19_cast_fp16, y = var_1300_to_fp16)[name = string("op_1301_cast_fp16")]; + tensor var_1302 = const()[name = string("op_1302"), val = tensor([1, 20, 64, -1])]; + tensor var_1303_cast_fp16 = reshape(shape = var_1302, x = key_19_cast_fp16)[name = string("op_1303_cast_fp16")]; + bool mh_w_19_transpose_x_0 = const()[name = string("mh_w_19_transpose_x_0"), val = bool(true)]; + bool mh_w_19_transpose_y_0 = const()[name = string("mh_w_19_transpose_y_0"), val = bool(false)]; + tensor mh_w_19_cast_fp16 = matmul(transpose_x = mh_w_19_transpose_x_0, transpose_y = mh_w_19_transpose_y_0, x = var_1301_cast_fp16, y = var_1303_cast_fp16)[name = string("mh_w_19_cast_fp16")]; + tensor var_1306_cast_fp16 = softmax(axis = var_1244, x = mh_w_19_cast_fp16)[name = string("op_1306_cast_fp16")]; + tensor var_1307 = const()[name = string("op_1307"), val = tensor([1, 20, 64, -1])]; + tensor var_1308_cast_fp16 = reshape(shape = var_1307, x = value_19_cast_fp16)[name = string("op_1308_cast_fp16")]; + bool attn_19_transpose_x_0 = const()[name = string("attn_19_transpose_x_0"), val = bool(false)]; + bool attn_19_transpose_y_0 = const()[name = string("attn_19_transpose_y_0"), val = bool(true)]; + tensor attn_19_cast_fp16 = matmul(transpose_x = attn_19_transpose_x_0, transpose_y = attn_19_transpose_y_0, x = var_1308_cast_fp16, y = var_1306_cast_fp16)[name = string("attn_19_cast_fp16")]; + tensor var_1311 = const()[name = string("op_1311"), val = tensor([1, 1280, 1, -1])]; + tensor input_73_cast_fp16 = reshape(shape = var_1311, x = attn_19_cast_fp16)[name = string("input_73_cast_fp16")]; + string obj_39_pad_type_0 = const()[name = string("obj_39_pad_type_0"), val = string("valid")]; + tensor obj_39_strides_0 = const()[name = string("obj_39_strides_0"), val = tensor([1, 1])]; + tensor obj_39_pad_0 = const()[name = string("obj_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_39_dilations_0 = const()[name = string("obj_39_dilations_0"), val = tensor([1, 1])]; + int32 obj_39_groups_0 = const()[name = string("obj_39_groups_0"), val = int32(1)]; + tensor layers_9_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(378684800)))]; + tensor layers_9_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_9_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381961664)))]; + tensor obj_39_cast_fp16 = conv(bias = layers_9_self_attn_o_proj_bias_to_fp16, dilations = obj_39_dilations_0, groups = obj_39_groups_0, pad = obj_39_pad_0, pad_type = obj_39_pad_type_0, strides = obj_39_strides_0, weight = layers_9_self_attn_o_proj_weight_to_fp16, x = input_73_cast_fp16)[name = string("obj_39_cast_fp16")]; + tensor inputs_39_cast_fp16 = add(x = inputs_37_cast_fp16, y = obj_39_cast_fp16)[name = string("inputs_39_cast_fp16")]; + tensor out_39_axes_0 = const()[name = string("out_39_axes_0"), val = tensor([1])]; + fp16 var_1329_to_fp16 = const()[name = string("op_1329_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_39_cast_fp16 = layer_norm(axes = out_39_axes_0, epsilon = var_1329_to_fp16, x = inputs_39_cast_fp16)[name = string("out_39_cast_fp16")]; + tensor input_75_gamma_0_to_fp16 = const()[name = string("input_75_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381964288)))]; + tensor input_75_beta_0_to_fp16 = const()[name = string("input_75_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381966912)))]; + fp16 input_75_epsilon_0_to_fp16 = const()[name = string("input_75_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor 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 = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_39_cast_fp16)[name = string("input_75_cast_fp16")]; + string input_77_pad_type_0 = const()[name = string("input_77_pad_type_0"), val = string("valid")]; + tensor input_77_strides_0 = const()[name = string("input_77_strides_0"), val = tensor([1, 1])]; + tensor input_77_pad_0 = const()[name = string("input_77_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_77_dilations_0 = const()[name = string("input_77_dilations_0"), val = tensor([1, 1])]; + int32 input_77_groups_0 = const()[name = string("input_77_groups_0"), val = int32(1)]; + tensor layers_9_fc1_weight_to_fp16 = const()[name = string("layers_9_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381969536)))]; + tensor layers_9_fc1_bias_to_fp16 = const()[name = string("layers_9_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395076800)))]; + tensor input_77_cast_fp16 = conv(bias = layers_9_fc1_bias_to_fp16, 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_9_fc1_weight_to_fp16, x = input_75_cast_fp16)[name = string("input_77_cast_fp16")]; + string input_79_mode_0 = const()[name = string("input_79_mode_0"), val = string("EXACT")]; + tensor input_79_cast_fp16 = gelu(mode = input_79_mode_0, x = input_77_cast_fp16)[name = string("input_79_cast_fp16")]; + string hidden_states_23_pad_type_0 = const()[name = string("hidden_states_23_pad_type_0"), val = string("valid")]; + tensor hidden_states_23_strides_0 = const()[name = string("hidden_states_23_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_23_pad_0 = const()[name = string("hidden_states_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_23_dilations_0 = const()[name = string("hidden_states_23_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_23_groups_0 = const()[name = string("hidden_states_23_groups_0"), val = int32(1)]; + tensor layers_9_fc2_weight_to_fp16 = const()[name = string("layers_9_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395087104)))]; + tensor layers_9_fc2_bias_to_fp16 = const()[name = string("layers_9_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408194368)))]; + tensor hidden_states_23_cast_fp16 = conv(bias = layers_9_fc2_bias_to_fp16, dilations = hidden_states_23_dilations_0, groups = hidden_states_23_groups_0, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = hidden_states_23_strides_0, weight = layers_9_fc2_weight_to_fp16, x = input_79_cast_fp16)[name = string("hidden_states_23_cast_fp16")]; + tensor inputs_41_cast_fp16 = add(x = inputs_39_cast_fp16, y = hidden_states_23_cast_fp16)[name = string("inputs_41_cast_fp16")]; + int32 var_1362 = const()[name = string("op_1362"), val = int32(3)]; + tensor out_41_axes_0 = const()[name = string("out_41_axes_0"), val = tensor([1])]; + fp16 var_1381_to_fp16 = const()[name = string("op_1381_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_41_cast_fp16 = layer_norm(axes = out_41_axes_0, epsilon = var_1381_to_fp16, x = inputs_41_cast_fp16)[name = string("out_41_cast_fp16")]; + tensor obj_41_gamma_0_to_fp16 = const()[name = string("obj_41_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408196992)))]; + tensor obj_41_beta_0_to_fp16 = const()[name = string("obj_41_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408199616)))]; + fp16 obj_41_epsilon_0_to_fp16 = const()[name = string("obj_41_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor obj_41_cast_fp16 = batch_norm(beta = obj_41_beta_0_to_fp16, epsilon = obj_41_epsilon_0_to_fp16, gamma = obj_41_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_41_cast_fp16)[name = string("obj_41_cast_fp16")]; + string query_21_pad_type_0 = const()[name = string("query_21_pad_type_0"), val = string("valid")]; + tensor query_21_strides_0 = const()[name = string("query_21_strides_0"), val = tensor([1, 1])]; + tensor query_21_pad_0 = const()[name = string("query_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_21_dilations_0 = const()[name = string("query_21_dilations_0"), val = tensor([1, 1])]; + int32 query_21_groups_0 = const()[name = string("query_21_groups_0"), val = int32(1)]; + tensor layers_10_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_10_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408202240)))]; + tensor layers_10_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_10_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411479104)))]; + tensor query_21_cast_fp16 = conv(bias = layers_10_self_attn_q_proj_bias_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_10_self_attn_q_proj_weight_to_fp16, x = obj_41_cast_fp16)[name = string("query_21_cast_fp16")]; + string key_21_pad_type_0 = const()[name = string("key_21_pad_type_0"), val = string("valid")]; + tensor key_21_strides_0 = const()[name = string("key_21_strides_0"), val = tensor([1, 1])]; + tensor key_21_pad_0 = const()[name = string("key_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_21_dilations_0 = const()[name = string("key_21_dilations_0"), val = tensor([1, 1])]; + int32 key_21_groups_0 = const()[name = string("key_21_groups_0"), val = int32(1)]; + tensor layers_10_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_10_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411481728)))]; + tensor key_21_cast_fp16 = conv(dilations = key_21_dilations_0, groups = key_21_groups_0, pad = key_21_pad_0, pad_type = key_21_pad_type_0, strides = key_21_strides_0, weight = layers_10_self_attn_k_proj_weight_to_fp16, x = obj_41_cast_fp16)[name = string("key_21_cast_fp16")]; + string value_21_pad_type_0 = const()[name = string("value_21_pad_type_0"), val = string("valid")]; + tensor value_21_strides_0 = const()[name = string("value_21_strides_0"), val = tensor([1, 1])]; + tensor value_21_pad_0 = const()[name = string("value_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_21_dilations_0 = const()[name = string("value_21_dilations_0"), val = tensor([1, 1])]; + int32 value_21_groups_0 = const()[name = string("value_21_groups_0"), val = int32(1)]; + tensor layers_10_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_10_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(414758592)))]; + tensor layers_10_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_10_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418035456)))]; + tensor value_21_cast_fp16 = conv(bias = layers_10_self_attn_v_proj_bias_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_41_cast_fp16)[name = string("value_21_cast_fp16")]; + tensor var_1416 = const()[name = string("op_1416"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_21_cast_fp16 = reshape(shape = var_1416, x = query_21_cast_fp16)[name = string("mh_q_21_cast_fp16")]; + fp16 var_1418_to_fp16 = const()[name = string("op_1418_to_fp16"), val = fp16(0x1p-3)]; + tensor var_1419_cast_fp16 = mul(x = mh_q_21_cast_fp16, y = var_1418_to_fp16)[name = string("op_1419_cast_fp16")]; + tensor var_1420 = const()[name = string("op_1420"), val = tensor([1, 20, 64, -1])]; + tensor var_1421_cast_fp16 = reshape(shape = var_1420, x = key_21_cast_fp16)[name = string("op_1421_cast_fp16")]; + bool mh_w_21_transpose_x_0 = const()[name = string("mh_w_21_transpose_x_0"), val = bool(true)]; + bool mh_w_21_transpose_y_0 = const()[name = string("mh_w_21_transpose_y_0"), val = bool(false)]; + tensor mh_w_21_cast_fp16 = matmul(transpose_x = mh_w_21_transpose_x_0, transpose_y = mh_w_21_transpose_y_0, x = var_1419_cast_fp16, y = var_1421_cast_fp16)[name = string("mh_w_21_cast_fp16")]; + tensor var_1424_cast_fp16 = softmax(axis = var_1362, x = mh_w_21_cast_fp16)[name = string("op_1424_cast_fp16")]; + tensor var_1425 = const()[name = string("op_1425"), val = tensor([1, 20, 64, -1])]; + tensor var_1426_cast_fp16 = reshape(shape = var_1425, x = value_21_cast_fp16)[name = string("op_1426_cast_fp16")]; + bool attn_21_transpose_x_0 = const()[name = string("attn_21_transpose_x_0"), val = bool(false)]; + bool attn_21_transpose_y_0 = const()[name = string("attn_21_transpose_y_0"), val = bool(true)]; + tensor attn_21_cast_fp16 = matmul(transpose_x = attn_21_transpose_x_0, transpose_y = attn_21_transpose_y_0, x = var_1426_cast_fp16, y = var_1424_cast_fp16)[name = string("attn_21_cast_fp16")]; + tensor var_1429 = const()[name = string("op_1429"), val = tensor([1, 1280, 1, -1])]; + tensor input_81_cast_fp16 = reshape(shape = var_1429, x = attn_21_cast_fp16)[name = string("input_81_cast_fp16")]; + string obj_43_pad_type_0 = const()[name = string("obj_43_pad_type_0"), val = string("valid")]; + tensor obj_43_strides_0 = const()[name = string("obj_43_strides_0"), val = tensor([1, 1])]; + tensor obj_43_pad_0 = const()[name = string("obj_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_43_dilations_0 = const()[name = string("obj_43_dilations_0"), val = tensor([1, 1])]; + int32 obj_43_groups_0 = const()[name = string("obj_43_groups_0"), val = int32(1)]; + tensor layers_10_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_10_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418038080)))]; + tensor layers_10_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_10_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421314944)))]; + tensor obj_43_cast_fp16 = conv(bias = layers_10_self_attn_o_proj_bias_to_fp16, dilations = obj_43_dilations_0, groups = obj_43_groups_0, pad = obj_43_pad_0, pad_type = obj_43_pad_type_0, strides = obj_43_strides_0, weight = layers_10_self_attn_o_proj_weight_to_fp16, x = input_81_cast_fp16)[name = string("obj_43_cast_fp16")]; + tensor inputs_43_cast_fp16 = add(x = inputs_41_cast_fp16, y = obj_43_cast_fp16)[name = string("inputs_43_cast_fp16")]; + tensor out_43_axes_0 = const()[name = string("out_43_axes_0"), val = tensor([1])]; + fp16 var_1447_to_fp16 = const()[name = string("op_1447_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_43_cast_fp16 = layer_norm(axes = out_43_axes_0, epsilon = var_1447_to_fp16, x = inputs_43_cast_fp16)[name = string("out_43_cast_fp16")]; + tensor input_83_gamma_0_to_fp16 = const()[name = string("input_83_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421317568)))]; + tensor input_83_beta_0_to_fp16 = const()[name = string("input_83_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421320192)))]; + fp16 input_83_epsilon_0_to_fp16 = const()[name = string("input_83_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor input_83_cast_fp16 = batch_norm(beta = input_83_beta_0_to_fp16, epsilon = input_83_epsilon_0_to_fp16, gamma = input_83_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_43_cast_fp16)[name = string("input_83_cast_fp16")]; + string input_85_pad_type_0 = const()[name = string("input_85_pad_type_0"), val = string("valid")]; + tensor input_85_strides_0 = const()[name = string("input_85_strides_0"), val = tensor([1, 1])]; + tensor input_85_pad_0 = const()[name = string("input_85_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_85_dilations_0 = const()[name = string("input_85_dilations_0"), val = tensor([1, 1])]; + int32 input_85_groups_0 = const()[name = string("input_85_groups_0"), val = int32(1)]; + tensor layers_10_fc1_weight_to_fp16 = const()[name = string("layers_10_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421322816)))]; + tensor layers_10_fc1_bias_to_fp16 = const()[name = string("layers_10_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(434430080)))]; + tensor input_85_cast_fp16 = conv(bias = layers_10_fc1_bias_to_fp16, dilations = input_85_dilations_0, groups = input_85_groups_0, pad = input_85_pad_0, pad_type = input_85_pad_type_0, strides = input_85_strides_0, weight = layers_10_fc1_weight_to_fp16, x = input_83_cast_fp16)[name = string("input_85_cast_fp16")]; + string input_87_mode_0 = const()[name = string("input_87_mode_0"), val = string("EXACT")]; + tensor input_87_cast_fp16 = gelu(mode = input_87_mode_0, x = input_85_cast_fp16)[name = string("input_87_cast_fp16")]; + string hidden_states_25_pad_type_0 = const()[name = string("hidden_states_25_pad_type_0"), val = string("valid")]; + tensor hidden_states_25_strides_0 = const()[name = string("hidden_states_25_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_25_pad_0 = const()[name = string("hidden_states_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_25_dilations_0 = const()[name = string("hidden_states_25_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_25_groups_0 = const()[name = string("hidden_states_25_groups_0"), val = int32(1)]; + tensor layers_10_fc2_weight_to_fp16 = const()[name = string("layers_10_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(434440384)))]; + tensor layers_10_fc2_bias_to_fp16 = const()[name = string("layers_10_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447547648)))]; + tensor hidden_states_25_cast_fp16 = conv(bias = layers_10_fc2_bias_to_fp16, dilations = hidden_states_25_dilations_0, groups = hidden_states_25_groups_0, pad = hidden_states_25_pad_0, pad_type = hidden_states_25_pad_type_0, strides = hidden_states_25_strides_0, weight = layers_10_fc2_weight_to_fp16, x = input_87_cast_fp16)[name = string("hidden_states_25_cast_fp16")]; + tensor inputs_45_cast_fp16 = add(x = inputs_43_cast_fp16, y = hidden_states_25_cast_fp16)[name = string("inputs_45_cast_fp16")]; + int32 var_1480 = const()[name = string("op_1480"), val = int32(3)]; + tensor out_45_axes_0 = const()[name = string("out_45_axes_0"), val = tensor([1])]; + fp16 var_1499_to_fp16 = const()[name = string("op_1499_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_45_cast_fp16 = layer_norm(axes = out_45_axes_0, epsilon = var_1499_to_fp16, x = inputs_45_cast_fp16)[name = string("out_45_cast_fp16")]; + tensor obj_45_gamma_0_to_fp16 = const()[name = string("obj_45_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447550272)))]; + tensor obj_45_beta_0_to_fp16 = const()[name = string("obj_45_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447552896)))]; + fp16 obj_45_epsilon_0_to_fp16 = const()[name = string("obj_45_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor obj_45_cast_fp16 = batch_norm(beta = obj_45_beta_0_to_fp16, epsilon = obj_45_epsilon_0_to_fp16, gamma = obj_45_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_45_cast_fp16)[name = string("obj_45_cast_fp16")]; + string query_23_pad_type_0 = const()[name = string("query_23_pad_type_0"), val = string("valid")]; + tensor query_23_strides_0 = const()[name = string("query_23_strides_0"), val = tensor([1, 1])]; + tensor query_23_pad_0 = const()[name = string("query_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_23_dilations_0 = const()[name = string("query_23_dilations_0"), val = tensor([1, 1])]; + int32 query_23_groups_0 = const()[name = string("query_23_groups_0"), val = int32(1)]; + tensor layers_11_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_11_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447555520)))]; + tensor layers_11_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_11_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450832384)))]; + tensor query_23_cast_fp16 = conv(bias = layers_11_self_attn_q_proj_bias_to_fp16, dilations = query_23_dilations_0, groups = query_23_groups_0, pad = query_23_pad_0, pad_type = query_23_pad_type_0, strides = query_23_strides_0, weight = layers_11_self_attn_q_proj_weight_to_fp16, x = obj_45_cast_fp16)[name = string("query_23_cast_fp16")]; + string key_23_pad_type_0 = const()[name = string("key_23_pad_type_0"), val = string("valid")]; + tensor key_23_strides_0 = const()[name = string("key_23_strides_0"), val = tensor([1, 1])]; + tensor key_23_pad_0 = const()[name = string("key_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_23_dilations_0 = const()[name = string("key_23_dilations_0"), val = tensor([1, 1])]; + int32 key_23_groups_0 = const()[name = string("key_23_groups_0"), val = int32(1)]; + tensor layers_11_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_11_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450835008)))]; + tensor key_23_cast_fp16 = conv(dilations = key_23_dilations_0, groups = key_23_groups_0, pad = key_23_pad_0, pad_type = key_23_pad_type_0, strides = key_23_strides_0, weight = layers_11_self_attn_k_proj_weight_to_fp16, x = obj_45_cast_fp16)[name = string("key_23_cast_fp16")]; + string value_23_pad_type_0 = const()[name = string("value_23_pad_type_0"), val = string("valid")]; + tensor value_23_strides_0 = const()[name = string("value_23_strides_0"), val = tensor([1, 1])]; + tensor value_23_pad_0 = const()[name = string("value_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_23_dilations_0 = const()[name = string("value_23_dilations_0"), val = tensor([1, 1])]; + int32 value_23_groups_0 = const()[name = string("value_23_groups_0"), val = int32(1)]; + tensor layers_11_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_11_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454111872)))]; + tensor layers_11_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_11_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457388736)))]; + tensor value_23_cast_fp16 = conv(bias = layers_11_self_attn_v_proj_bias_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_45_cast_fp16)[name = string("value_23_cast_fp16")]; + tensor var_1534 = const()[name = string("op_1534"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_23_cast_fp16 = reshape(shape = var_1534, x = query_23_cast_fp16)[name = string("mh_q_23_cast_fp16")]; + fp16 var_1536_to_fp16 = const()[name = string("op_1536_to_fp16"), val = fp16(0x1p-3)]; + tensor var_1537_cast_fp16 = mul(x = mh_q_23_cast_fp16, y = var_1536_to_fp16)[name = string("op_1537_cast_fp16")]; + tensor var_1538 = const()[name = string("op_1538"), val = tensor([1, 20, 64, -1])]; + tensor var_1539_cast_fp16 = reshape(shape = var_1538, x = key_23_cast_fp16)[name = string("op_1539_cast_fp16")]; + bool mh_w_23_transpose_x_0 = const()[name = string("mh_w_23_transpose_x_0"), val = bool(true)]; + bool mh_w_23_transpose_y_0 = const()[name = string("mh_w_23_transpose_y_0"), val = bool(false)]; + tensor mh_w_23_cast_fp16 = matmul(transpose_x = mh_w_23_transpose_x_0, transpose_y = mh_w_23_transpose_y_0, x = var_1537_cast_fp16, y = var_1539_cast_fp16)[name = string("mh_w_23_cast_fp16")]; + tensor var_1542_cast_fp16 = softmax(axis = var_1480, x = mh_w_23_cast_fp16)[name = string("op_1542_cast_fp16")]; + tensor var_1543 = const()[name = string("op_1543"), val = tensor([1, 20, 64, -1])]; + tensor var_1544_cast_fp16 = reshape(shape = var_1543, x = value_23_cast_fp16)[name = string("op_1544_cast_fp16")]; + bool attn_23_transpose_x_0 = const()[name = string("attn_23_transpose_x_0"), val = bool(false)]; + bool attn_23_transpose_y_0 = const()[name = string("attn_23_transpose_y_0"), val = bool(true)]; + tensor attn_23_cast_fp16 = matmul(transpose_x = attn_23_transpose_x_0, transpose_y = attn_23_transpose_y_0, x = var_1544_cast_fp16, y = var_1542_cast_fp16)[name = string("attn_23_cast_fp16")]; + tensor var_1547 = const()[name = string("op_1547"), val = tensor([1, 1280, 1, -1])]; + tensor input_89_cast_fp16 = reshape(shape = var_1547, x = attn_23_cast_fp16)[name = string("input_89_cast_fp16")]; + string obj_47_pad_type_0 = const()[name = string("obj_47_pad_type_0"), val = string("valid")]; + tensor obj_47_strides_0 = const()[name = string("obj_47_strides_0"), val = tensor([1, 1])]; + tensor obj_47_pad_0 = const()[name = string("obj_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_47_dilations_0 = const()[name = string("obj_47_dilations_0"), val = tensor([1, 1])]; + int32 obj_47_groups_0 = const()[name = string("obj_47_groups_0"), val = int32(1)]; + tensor layers_11_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_11_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457391360)))]; + tensor layers_11_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_11_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(460668224)))]; + tensor obj_47_cast_fp16 = conv(bias = layers_11_self_attn_o_proj_bias_to_fp16, dilations = obj_47_dilations_0, groups = obj_47_groups_0, pad = obj_47_pad_0, pad_type = obj_47_pad_type_0, strides = obj_47_strides_0, weight = layers_11_self_attn_o_proj_weight_to_fp16, x = input_89_cast_fp16)[name = string("obj_47_cast_fp16")]; + tensor inputs_47_cast_fp16 = add(x = inputs_45_cast_fp16, y = obj_47_cast_fp16)[name = string("inputs_47_cast_fp16")]; + tensor out_47_axes_0 = const()[name = string("out_47_axes_0"), val = tensor([1])]; + fp16 var_1565_to_fp16 = const()[name = string("op_1565_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_47_cast_fp16 = layer_norm(axes = out_47_axes_0, epsilon = var_1565_to_fp16, x = inputs_47_cast_fp16)[name = string("out_47_cast_fp16")]; + tensor input_91_gamma_0_to_fp16 = const()[name = string("input_91_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(460670848)))]; + tensor input_91_beta_0_to_fp16 = const()[name = string("input_91_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(460673472)))]; + fp16 input_91_epsilon_0_to_fp16 = const()[name = string("input_91_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor input_91_cast_fp16 = batch_norm(beta = input_91_beta_0_to_fp16, epsilon = input_91_epsilon_0_to_fp16, gamma = input_91_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_47_cast_fp16)[name = string("input_91_cast_fp16")]; + string input_93_pad_type_0 = const()[name = string("input_93_pad_type_0"), val = string("valid")]; + tensor input_93_strides_0 = const()[name = string("input_93_strides_0"), val = tensor([1, 1])]; + tensor input_93_pad_0 = const()[name = string("input_93_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_93_dilations_0 = const()[name = string("input_93_dilations_0"), val = tensor([1, 1])]; + int32 input_93_groups_0 = const()[name = string("input_93_groups_0"), val = int32(1)]; + tensor layers_11_fc1_weight_to_fp16 = const()[name = string("layers_11_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(460676096)))]; + tensor layers_11_fc1_bias_to_fp16 = const()[name = string("layers_11_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473783360)))]; + tensor input_93_cast_fp16 = conv(bias = layers_11_fc1_bias_to_fp16, dilations = input_93_dilations_0, groups = input_93_groups_0, pad = input_93_pad_0, pad_type = input_93_pad_type_0, strides = input_93_strides_0, weight = layers_11_fc1_weight_to_fp16, x = input_91_cast_fp16)[name = string("input_93_cast_fp16")]; + string input_95_mode_0 = const()[name = string("input_95_mode_0"), val = string("EXACT")]; + tensor input_95_cast_fp16 = gelu(mode = input_95_mode_0, x = input_93_cast_fp16)[name = string("input_95_cast_fp16")]; + string hidden_states_27_pad_type_0 = const()[name = string("hidden_states_27_pad_type_0"), val = string("valid")]; + tensor hidden_states_27_strides_0 = const()[name = string("hidden_states_27_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_27_pad_0 = const()[name = string("hidden_states_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_27_dilations_0 = const()[name = string("hidden_states_27_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_27_groups_0 = const()[name = string("hidden_states_27_groups_0"), val = int32(1)]; + tensor layers_11_fc2_weight_to_fp16 = const()[name = string("layers_11_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473793664)))]; + tensor layers_11_fc2_bias_to_fp16 = const()[name = string("layers_11_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486900928)))]; + tensor hidden_states_27_cast_fp16 = conv(bias = layers_11_fc2_bias_to_fp16, dilations = hidden_states_27_dilations_0, groups = hidden_states_27_groups_0, pad = hidden_states_27_pad_0, pad_type = hidden_states_27_pad_type_0, strides = hidden_states_27_strides_0, weight = layers_11_fc2_weight_to_fp16, x = input_95_cast_fp16)[name = string("hidden_states_27_cast_fp16")]; + tensor inputs_49_cast_fp16 = add(x = inputs_47_cast_fp16, y = hidden_states_27_cast_fp16)[name = string("inputs_49_cast_fp16")]; + int32 var_1598 = const()[name = string("op_1598"), val = int32(3)]; + tensor out_49_axes_0 = const()[name = string("out_49_axes_0"), val = tensor([1])]; + fp16 var_1617_to_fp16 = const()[name = string("op_1617_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_49_cast_fp16 = layer_norm(axes = out_49_axes_0, epsilon = var_1617_to_fp16, x = inputs_49_cast_fp16)[name = string("out_49_cast_fp16")]; + tensor obj_49_gamma_0_to_fp16 = const()[name = string("obj_49_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486903552)))]; + tensor obj_49_beta_0_to_fp16 = const()[name = string("obj_49_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486906176)))]; + fp16 obj_49_epsilon_0_to_fp16 = const()[name = string("obj_49_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor obj_49_cast_fp16 = batch_norm(beta = obj_49_beta_0_to_fp16, epsilon = obj_49_epsilon_0_to_fp16, gamma = obj_49_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_49_cast_fp16)[name = string("obj_49_cast_fp16")]; + string query_25_pad_type_0 = const()[name = string("query_25_pad_type_0"), val = string("valid")]; + tensor query_25_strides_0 = const()[name = string("query_25_strides_0"), val = tensor([1, 1])]; + tensor query_25_pad_0 = const()[name = string("query_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_25_dilations_0 = const()[name = string("query_25_dilations_0"), val = tensor([1, 1])]; + int32 query_25_groups_0 = const()[name = string("query_25_groups_0"), val = int32(1)]; + tensor layers_12_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_12_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486908800)))]; + tensor layers_12_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_12_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490185664)))]; + tensor query_25_cast_fp16 = conv(bias = layers_12_self_attn_q_proj_bias_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_12_self_attn_q_proj_weight_to_fp16, x = obj_49_cast_fp16)[name = string("query_25_cast_fp16")]; + string key_25_pad_type_0 = const()[name = string("key_25_pad_type_0"), val = string("valid")]; + tensor key_25_strides_0 = const()[name = string("key_25_strides_0"), val = tensor([1, 1])]; + tensor key_25_pad_0 = const()[name = string("key_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_25_dilations_0 = const()[name = string("key_25_dilations_0"), val = tensor([1, 1])]; + int32 key_25_groups_0 = const()[name = string("key_25_groups_0"), val = int32(1)]; + tensor layers_12_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_12_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490188288)))]; + tensor key_25_cast_fp16 = conv(dilations = key_25_dilations_0, groups = key_25_groups_0, pad = key_25_pad_0, pad_type = key_25_pad_type_0, strides = key_25_strides_0, weight = layers_12_self_attn_k_proj_weight_to_fp16, x = obj_49_cast_fp16)[name = string("key_25_cast_fp16")]; + string value_25_pad_type_0 = const()[name = string("value_25_pad_type_0"), val = string("valid")]; + tensor value_25_strides_0 = const()[name = string("value_25_strides_0"), val = tensor([1, 1])]; + tensor value_25_pad_0 = const()[name = string("value_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_25_dilations_0 = const()[name = string("value_25_dilations_0"), val = tensor([1, 1])]; + int32 value_25_groups_0 = const()[name = string("value_25_groups_0"), val = int32(1)]; + tensor layers_12_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_12_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493465152)))]; + tensor layers_12_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_12_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(496742016)))]; + tensor value_25_cast_fp16 = conv(bias = layers_12_self_attn_v_proj_bias_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_49_cast_fp16)[name = string("value_25_cast_fp16")]; + tensor var_1652 = const()[name = string("op_1652"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_25_cast_fp16 = reshape(shape = var_1652, x = query_25_cast_fp16)[name = string("mh_q_25_cast_fp16")]; + fp16 var_1654_to_fp16 = const()[name = string("op_1654_to_fp16"), val = fp16(0x1p-3)]; + tensor var_1655_cast_fp16 = mul(x = mh_q_25_cast_fp16, y = var_1654_to_fp16)[name = string("op_1655_cast_fp16")]; + tensor var_1656 = const()[name = string("op_1656"), val = tensor([1, 20, 64, -1])]; + tensor var_1657_cast_fp16 = reshape(shape = var_1656, x = key_25_cast_fp16)[name = string("op_1657_cast_fp16")]; + bool mh_w_25_transpose_x_0 = const()[name = string("mh_w_25_transpose_x_0"), val = bool(true)]; + bool mh_w_25_transpose_y_0 = const()[name = string("mh_w_25_transpose_y_0"), val = bool(false)]; + tensor mh_w_25_cast_fp16 = matmul(transpose_x = mh_w_25_transpose_x_0, transpose_y = mh_w_25_transpose_y_0, x = var_1655_cast_fp16, y = var_1657_cast_fp16)[name = string("mh_w_25_cast_fp16")]; + tensor var_1660_cast_fp16 = softmax(axis = var_1598, x = mh_w_25_cast_fp16)[name = string("op_1660_cast_fp16")]; + tensor var_1661 = const()[name = string("op_1661"), val = tensor([1, 20, 64, -1])]; + tensor var_1662_cast_fp16 = reshape(shape = var_1661, x = value_25_cast_fp16)[name = string("op_1662_cast_fp16")]; + bool attn_25_transpose_x_0 = const()[name = string("attn_25_transpose_x_0"), val = bool(false)]; + bool attn_25_transpose_y_0 = const()[name = string("attn_25_transpose_y_0"), val = bool(true)]; + tensor attn_25_cast_fp16 = matmul(transpose_x = attn_25_transpose_x_0, transpose_y = attn_25_transpose_y_0, x = var_1662_cast_fp16, y = var_1660_cast_fp16)[name = string("attn_25_cast_fp16")]; + tensor var_1665 = const()[name = string("op_1665"), val = tensor([1, 1280, 1, -1])]; + tensor input_97_cast_fp16 = reshape(shape = var_1665, x = attn_25_cast_fp16)[name = string("input_97_cast_fp16")]; + string obj_51_pad_type_0 = const()[name = string("obj_51_pad_type_0"), val = string("valid")]; + tensor obj_51_strides_0 = const()[name = string("obj_51_strides_0"), val = tensor([1, 1])]; + tensor obj_51_pad_0 = const()[name = string("obj_51_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_51_dilations_0 = const()[name = string("obj_51_dilations_0"), val = tensor([1, 1])]; + int32 obj_51_groups_0 = const()[name = string("obj_51_groups_0"), val = int32(1)]; + tensor layers_12_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_12_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(496744640)))]; + tensor layers_12_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_12_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(500021504)))]; + tensor obj_51_cast_fp16 = conv(bias = layers_12_self_attn_o_proj_bias_to_fp16, dilations = obj_51_dilations_0, groups = obj_51_groups_0, pad = obj_51_pad_0, pad_type = obj_51_pad_type_0, strides = obj_51_strides_0, weight = layers_12_self_attn_o_proj_weight_to_fp16, x = input_97_cast_fp16)[name = string("obj_51_cast_fp16")]; + tensor inputs_51_cast_fp16 = add(x = inputs_49_cast_fp16, y = obj_51_cast_fp16)[name = string("inputs_51_cast_fp16")]; + tensor out_51_axes_0 = const()[name = string("out_51_axes_0"), val = tensor([1])]; + fp16 var_1683_to_fp16 = const()[name = string("op_1683_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_51_cast_fp16 = layer_norm(axes = out_51_axes_0, epsilon = var_1683_to_fp16, x = inputs_51_cast_fp16)[name = string("out_51_cast_fp16")]; + tensor input_99_gamma_0_to_fp16 = const()[name = string("input_99_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(500024128)))]; + tensor input_99_beta_0_to_fp16 = const()[name = string("input_99_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(500026752)))]; + fp16 input_99_epsilon_0_to_fp16 = const()[name = string("input_99_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor input_99_cast_fp16 = batch_norm(beta = input_99_beta_0_to_fp16, epsilon = input_99_epsilon_0_to_fp16, gamma = input_99_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_51_cast_fp16)[name = string("input_99_cast_fp16")]; + string input_101_pad_type_0 = const()[name = string("input_101_pad_type_0"), val = string("valid")]; + tensor input_101_strides_0 = const()[name = string("input_101_strides_0"), val = tensor([1, 1])]; + tensor input_101_pad_0 = const()[name = string("input_101_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_101_dilations_0 = const()[name = string("input_101_dilations_0"), val = tensor([1, 1])]; + int32 input_101_groups_0 = const()[name = string("input_101_groups_0"), val = int32(1)]; + tensor layers_12_fc1_weight_to_fp16 = const()[name = string("layers_12_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(500029376)))]; + tensor layers_12_fc1_bias_to_fp16 = const()[name = string("layers_12_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513136640)))]; + tensor input_101_cast_fp16 = conv(bias = layers_12_fc1_bias_to_fp16, dilations = input_101_dilations_0, groups = input_101_groups_0, pad = input_101_pad_0, pad_type = input_101_pad_type_0, strides = input_101_strides_0, weight = layers_12_fc1_weight_to_fp16, x = input_99_cast_fp16)[name = string("input_101_cast_fp16")]; + string input_103_mode_0 = const()[name = string("input_103_mode_0"), val = string("EXACT")]; + tensor input_103_cast_fp16 = gelu(mode = input_103_mode_0, x = input_101_cast_fp16)[name = string("input_103_cast_fp16")]; + string hidden_states_29_pad_type_0 = const()[name = string("hidden_states_29_pad_type_0"), val = string("valid")]; + tensor hidden_states_29_strides_0 = const()[name = string("hidden_states_29_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_29_pad_0 = const()[name = string("hidden_states_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_29_dilations_0 = const()[name = string("hidden_states_29_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_29_groups_0 = const()[name = string("hidden_states_29_groups_0"), val = int32(1)]; + tensor layers_12_fc2_weight_to_fp16 = const()[name = string("layers_12_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513146944)))]; + tensor layers_12_fc2_bias_to_fp16 = const()[name = string("layers_12_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526254208)))]; + tensor hidden_states_29_cast_fp16 = conv(bias = layers_12_fc2_bias_to_fp16, dilations = hidden_states_29_dilations_0, groups = hidden_states_29_groups_0, pad = hidden_states_29_pad_0, pad_type = hidden_states_29_pad_type_0, strides = hidden_states_29_strides_0, weight = layers_12_fc2_weight_to_fp16, x = input_103_cast_fp16)[name = string("hidden_states_29_cast_fp16")]; + tensor inputs_53_cast_fp16 = add(x = inputs_51_cast_fp16, y = hidden_states_29_cast_fp16)[name = string("inputs_53_cast_fp16")]; + int32 var_1716 = const()[name = string("op_1716"), val = int32(3)]; + tensor out_53_axes_0 = const()[name = string("out_53_axes_0"), val = tensor([1])]; + fp16 var_1735_to_fp16 = const()[name = string("op_1735_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_53_cast_fp16 = layer_norm(axes = out_53_axes_0, epsilon = var_1735_to_fp16, x = inputs_53_cast_fp16)[name = string("out_53_cast_fp16")]; + tensor obj_53_gamma_0_to_fp16 = const()[name = string("obj_53_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526256832)))]; + tensor obj_53_beta_0_to_fp16 = const()[name = string("obj_53_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526259456)))]; + fp16 obj_53_epsilon_0_to_fp16 = const()[name = string("obj_53_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor obj_53_cast_fp16 = batch_norm(beta = obj_53_beta_0_to_fp16, epsilon = obj_53_epsilon_0_to_fp16, gamma = obj_53_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_53_cast_fp16)[name = string("obj_53_cast_fp16")]; + string query_27_pad_type_0 = const()[name = string("query_27_pad_type_0"), val = string("valid")]; + tensor query_27_strides_0 = const()[name = string("query_27_strides_0"), val = tensor([1, 1])]; + tensor query_27_pad_0 = const()[name = string("query_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_27_dilations_0 = const()[name = string("query_27_dilations_0"), val = tensor([1, 1])]; + int32 query_27_groups_0 = const()[name = string("query_27_groups_0"), val = int32(1)]; + tensor layers_13_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_13_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526262080)))]; + tensor layers_13_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_13_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529538944)))]; + tensor query_27_cast_fp16 = conv(bias = layers_13_self_attn_q_proj_bias_to_fp16, dilations = query_27_dilations_0, groups = query_27_groups_0, pad = query_27_pad_0, pad_type = query_27_pad_type_0, strides = query_27_strides_0, weight = layers_13_self_attn_q_proj_weight_to_fp16, x = obj_53_cast_fp16)[name = string("query_27_cast_fp16")]; + string key_27_pad_type_0 = const()[name = string("key_27_pad_type_0"), val = string("valid")]; + tensor key_27_strides_0 = const()[name = string("key_27_strides_0"), val = tensor([1, 1])]; + tensor key_27_pad_0 = const()[name = string("key_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_27_dilations_0 = const()[name = string("key_27_dilations_0"), val = tensor([1, 1])]; + int32 key_27_groups_0 = const()[name = string("key_27_groups_0"), val = int32(1)]; + tensor layers_13_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_13_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529541568)))]; + tensor key_27_cast_fp16 = conv(dilations = key_27_dilations_0, groups = key_27_groups_0, pad = key_27_pad_0, pad_type = key_27_pad_type_0, strides = key_27_strides_0, weight = layers_13_self_attn_k_proj_weight_to_fp16, x = obj_53_cast_fp16)[name = string("key_27_cast_fp16")]; + string value_27_pad_type_0 = const()[name = string("value_27_pad_type_0"), val = string("valid")]; + tensor value_27_strides_0 = const()[name = string("value_27_strides_0"), val = tensor([1, 1])]; + tensor value_27_pad_0 = const()[name = string("value_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_27_dilations_0 = const()[name = string("value_27_dilations_0"), val = tensor([1, 1])]; + int32 value_27_groups_0 = const()[name = string("value_27_groups_0"), val = int32(1)]; + tensor layers_13_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_13_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(532818432)))]; + tensor layers_13_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_13_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(536095296)))]; + tensor value_27_cast_fp16 = conv(bias = layers_13_self_attn_v_proj_bias_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_53_cast_fp16)[name = string("value_27_cast_fp16")]; + tensor var_1770 = const()[name = string("op_1770"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_27_cast_fp16 = reshape(shape = var_1770, x = query_27_cast_fp16)[name = string("mh_q_27_cast_fp16")]; + fp16 var_1772_to_fp16 = const()[name = string("op_1772_to_fp16"), val = fp16(0x1p-3)]; + tensor var_1773_cast_fp16 = mul(x = mh_q_27_cast_fp16, y = var_1772_to_fp16)[name = string("op_1773_cast_fp16")]; + tensor var_1774 = const()[name = string("op_1774"), val = tensor([1, 20, 64, -1])]; + tensor var_1775_cast_fp16 = reshape(shape = var_1774, x = key_27_cast_fp16)[name = string("op_1775_cast_fp16")]; + bool mh_w_27_transpose_x_0 = const()[name = string("mh_w_27_transpose_x_0"), val = bool(true)]; + bool mh_w_27_transpose_y_0 = const()[name = string("mh_w_27_transpose_y_0"), val = bool(false)]; + tensor mh_w_27_cast_fp16 = matmul(transpose_x = mh_w_27_transpose_x_0, transpose_y = mh_w_27_transpose_y_0, x = var_1773_cast_fp16, y = var_1775_cast_fp16)[name = string("mh_w_27_cast_fp16")]; + tensor var_1778_cast_fp16 = softmax(axis = var_1716, x = mh_w_27_cast_fp16)[name = string("op_1778_cast_fp16")]; + tensor var_1779 = const()[name = string("op_1779"), val = tensor([1, 20, 64, -1])]; + tensor var_1780_cast_fp16 = reshape(shape = var_1779, x = value_27_cast_fp16)[name = string("op_1780_cast_fp16")]; + bool attn_27_transpose_x_0 = const()[name = string("attn_27_transpose_x_0"), val = bool(false)]; + bool attn_27_transpose_y_0 = const()[name = string("attn_27_transpose_y_0"), val = bool(true)]; + tensor attn_27_cast_fp16 = matmul(transpose_x = attn_27_transpose_x_0, transpose_y = attn_27_transpose_y_0, x = var_1780_cast_fp16, y = var_1778_cast_fp16)[name = string("attn_27_cast_fp16")]; + tensor var_1783 = const()[name = string("op_1783"), val = tensor([1, 1280, 1, -1])]; + tensor input_105_cast_fp16 = reshape(shape = var_1783, x = attn_27_cast_fp16)[name = string("input_105_cast_fp16")]; + string obj_55_pad_type_0 = const()[name = string("obj_55_pad_type_0"), val = string("valid")]; + tensor obj_55_strides_0 = const()[name = string("obj_55_strides_0"), val = tensor([1, 1])]; + tensor obj_55_pad_0 = const()[name = string("obj_55_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_55_dilations_0 = const()[name = string("obj_55_dilations_0"), val = tensor([1, 1])]; + int32 obj_55_groups_0 = const()[name = string("obj_55_groups_0"), val = int32(1)]; + tensor layers_13_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_13_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(536097920)))]; + tensor layers_13_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_13_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(539374784)))]; + tensor obj_55_cast_fp16 = conv(bias = layers_13_self_attn_o_proj_bias_to_fp16, dilations = obj_55_dilations_0, groups = obj_55_groups_0, pad = obj_55_pad_0, pad_type = obj_55_pad_type_0, strides = obj_55_strides_0, weight = layers_13_self_attn_o_proj_weight_to_fp16, x = input_105_cast_fp16)[name = string("obj_55_cast_fp16")]; + tensor inputs_55_cast_fp16 = add(x = inputs_53_cast_fp16, y = obj_55_cast_fp16)[name = string("inputs_55_cast_fp16")]; + tensor out_55_axes_0 = const()[name = string("out_55_axes_0"), val = tensor([1])]; + fp16 var_1801_to_fp16 = const()[name = string("op_1801_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_55_cast_fp16 = layer_norm(axes = out_55_axes_0, epsilon = var_1801_to_fp16, x = inputs_55_cast_fp16)[name = string("out_55_cast_fp16")]; + tensor input_107_gamma_0_to_fp16 = const()[name = string("input_107_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(539377408)))]; + tensor input_107_beta_0_to_fp16 = const()[name = string("input_107_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(539380032)))]; + fp16 input_107_epsilon_0_to_fp16 = const()[name = string("input_107_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor input_107_cast_fp16 = batch_norm(beta = input_107_beta_0_to_fp16, epsilon = input_107_epsilon_0_to_fp16, gamma = input_107_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_55_cast_fp16)[name = string("input_107_cast_fp16")]; + string input_109_pad_type_0 = const()[name = string("input_109_pad_type_0"), val = string("valid")]; + tensor input_109_strides_0 = const()[name = string("input_109_strides_0"), val = tensor([1, 1])]; + tensor input_109_pad_0 = const()[name = string("input_109_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_109_dilations_0 = const()[name = string("input_109_dilations_0"), val = tensor([1, 1])]; + int32 input_109_groups_0 = const()[name = string("input_109_groups_0"), val = int32(1)]; + tensor layers_13_fc1_weight_to_fp16 = const()[name = string("layers_13_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(539382656)))]; + tensor layers_13_fc1_bias_to_fp16 = const()[name = string("layers_13_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(552489920)))]; + tensor input_109_cast_fp16 = conv(bias = layers_13_fc1_bias_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 = layers_13_fc1_weight_to_fp16, x = input_107_cast_fp16)[name = string("input_109_cast_fp16")]; + string input_111_mode_0 = const()[name = string("input_111_mode_0"), val = string("EXACT")]; + tensor input_111_cast_fp16 = gelu(mode = input_111_mode_0, x = input_109_cast_fp16)[name = string("input_111_cast_fp16")]; + string hidden_states_31_pad_type_0 = const()[name = string("hidden_states_31_pad_type_0"), val = string("valid")]; + tensor hidden_states_31_strides_0 = const()[name = string("hidden_states_31_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_31_pad_0 = const()[name = string("hidden_states_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_31_dilations_0 = const()[name = string("hidden_states_31_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_31_groups_0 = const()[name = string("hidden_states_31_groups_0"), val = int32(1)]; + tensor layers_13_fc2_weight_to_fp16 = const()[name = string("layers_13_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(552500224)))]; + tensor layers_13_fc2_bias_to_fp16 = const()[name = string("layers_13_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(565607488)))]; + tensor hidden_states_31_cast_fp16 = conv(bias = layers_13_fc2_bias_to_fp16, dilations = hidden_states_31_dilations_0, groups = hidden_states_31_groups_0, pad = hidden_states_31_pad_0, pad_type = hidden_states_31_pad_type_0, strides = hidden_states_31_strides_0, weight = layers_13_fc2_weight_to_fp16, x = input_111_cast_fp16)[name = string("hidden_states_31_cast_fp16")]; + tensor inputs_57_cast_fp16 = add(x = inputs_55_cast_fp16, y = hidden_states_31_cast_fp16)[name = string("inputs_57_cast_fp16")]; + int32 var_1834 = const()[name = string("op_1834"), val = int32(3)]; + tensor out_57_axes_0 = const()[name = string("out_57_axes_0"), val = tensor([1])]; + fp16 var_1853_to_fp16 = const()[name = string("op_1853_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_57_cast_fp16 = layer_norm(axes = out_57_axes_0, epsilon = var_1853_to_fp16, x = inputs_57_cast_fp16)[name = string("out_57_cast_fp16")]; + tensor obj_57_gamma_0_to_fp16 = const()[name = string("obj_57_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(565610112)))]; + tensor obj_57_beta_0_to_fp16 = const()[name = string("obj_57_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(565612736)))]; + fp16 obj_57_epsilon_0_to_fp16 = const()[name = string("obj_57_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor obj_57_cast_fp16 = batch_norm(beta = obj_57_beta_0_to_fp16, epsilon = obj_57_epsilon_0_to_fp16, gamma = obj_57_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_57_cast_fp16)[name = string("obj_57_cast_fp16")]; + string query_29_pad_type_0 = const()[name = string("query_29_pad_type_0"), val = string("valid")]; + tensor query_29_strides_0 = const()[name = string("query_29_strides_0"), val = tensor([1, 1])]; + tensor query_29_pad_0 = const()[name = string("query_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_29_dilations_0 = const()[name = string("query_29_dilations_0"), val = tensor([1, 1])]; + int32 query_29_groups_0 = const()[name = string("query_29_groups_0"), val = int32(1)]; + tensor layers_14_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_14_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(565615360)))]; + tensor layers_14_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_14_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(568892224)))]; + tensor query_29_cast_fp16 = conv(bias = layers_14_self_attn_q_proj_bias_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_14_self_attn_q_proj_weight_to_fp16, x = obj_57_cast_fp16)[name = string("query_29_cast_fp16")]; + string key_29_pad_type_0 = const()[name = string("key_29_pad_type_0"), val = string("valid")]; + tensor key_29_strides_0 = const()[name = string("key_29_strides_0"), val = tensor([1, 1])]; + tensor key_29_pad_0 = const()[name = string("key_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_29_dilations_0 = const()[name = string("key_29_dilations_0"), val = tensor([1, 1])]; + int32 key_29_groups_0 = const()[name = string("key_29_groups_0"), val = int32(1)]; + tensor layers_14_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_14_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(568894848)))]; + tensor key_29_cast_fp16 = conv(dilations = key_29_dilations_0, groups = key_29_groups_0, pad = key_29_pad_0, pad_type = key_29_pad_type_0, strides = key_29_strides_0, weight = layers_14_self_attn_k_proj_weight_to_fp16, x = obj_57_cast_fp16)[name = string("key_29_cast_fp16")]; + string value_29_pad_type_0 = const()[name = string("value_29_pad_type_0"), val = string("valid")]; + tensor value_29_strides_0 = const()[name = string("value_29_strides_0"), val = tensor([1, 1])]; + tensor value_29_pad_0 = const()[name = string("value_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_29_dilations_0 = const()[name = string("value_29_dilations_0"), val = tensor([1, 1])]; + int32 value_29_groups_0 = const()[name = string("value_29_groups_0"), val = int32(1)]; + tensor layers_14_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_14_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572171712)))]; + tensor layers_14_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_14_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(575448576)))]; + tensor value_29_cast_fp16 = conv(bias = layers_14_self_attn_v_proj_bias_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_57_cast_fp16)[name = string("value_29_cast_fp16")]; + tensor var_1888 = const()[name = string("op_1888"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_29_cast_fp16 = reshape(shape = var_1888, x = query_29_cast_fp16)[name = string("mh_q_29_cast_fp16")]; + fp16 var_1890_to_fp16 = const()[name = string("op_1890_to_fp16"), val = fp16(0x1p-3)]; + tensor var_1891_cast_fp16 = mul(x = mh_q_29_cast_fp16, y = var_1890_to_fp16)[name = string("op_1891_cast_fp16")]; + tensor var_1892 = const()[name = string("op_1892"), val = tensor([1, 20, 64, -1])]; + tensor var_1893_cast_fp16 = reshape(shape = var_1892, x = key_29_cast_fp16)[name = string("op_1893_cast_fp16")]; + bool mh_w_29_transpose_x_0 = const()[name = string("mh_w_29_transpose_x_0"), val = bool(true)]; + bool mh_w_29_transpose_y_0 = const()[name = string("mh_w_29_transpose_y_0"), val = bool(false)]; + tensor mh_w_29_cast_fp16 = matmul(transpose_x = mh_w_29_transpose_x_0, transpose_y = mh_w_29_transpose_y_0, x = var_1891_cast_fp16, y = var_1893_cast_fp16)[name = string("mh_w_29_cast_fp16")]; + tensor var_1896_cast_fp16 = softmax(axis = var_1834, x = mh_w_29_cast_fp16)[name = string("op_1896_cast_fp16")]; + tensor var_1897 = const()[name = string("op_1897"), val = tensor([1, 20, 64, -1])]; + tensor var_1898_cast_fp16 = reshape(shape = var_1897, x = value_29_cast_fp16)[name = string("op_1898_cast_fp16")]; + bool attn_29_transpose_x_0 = const()[name = string("attn_29_transpose_x_0"), val = bool(false)]; + bool attn_29_transpose_y_0 = const()[name = string("attn_29_transpose_y_0"), val = bool(true)]; + tensor attn_29_cast_fp16 = matmul(transpose_x = attn_29_transpose_x_0, transpose_y = attn_29_transpose_y_0, x = var_1898_cast_fp16, y = var_1896_cast_fp16)[name = string("attn_29_cast_fp16")]; + tensor var_1901 = const()[name = string("op_1901"), val = tensor([1, 1280, 1, -1])]; + tensor input_113_cast_fp16 = reshape(shape = var_1901, x = attn_29_cast_fp16)[name = string("input_113_cast_fp16")]; + string obj_59_pad_type_0 = const()[name = string("obj_59_pad_type_0"), val = string("valid")]; + tensor obj_59_strides_0 = const()[name = string("obj_59_strides_0"), val = tensor([1, 1])]; + tensor obj_59_pad_0 = const()[name = string("obj_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_59_dilations_0 = const()[name = string("obj_59_dilations_0"), val = tensor([1, 1])]; + int32 obj_59_groups_0 = const()[name = string("obj_59_groups_0"), val = int32(1)]; + tensor layers_14_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_14_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(575451200)))]; + tensor layers_14_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_14_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(578728064)))]; + tensor obj_59_cast_fp16 = conv(bias = layers_14_self_attn_o_proj_bias_to_fp16, dilations = obj_59_dilations_0, groups = obj_59_groups_0, pad = obj_59_pad_0, pad_type = obj_59_pad_type_0, strides = obj_59_strides_0, weight = layers_14_self_attn_o_proj_weight_to_fp16, x = input_113_cast_fp16)[name = string("obj_59_cast_fp16")]; + tensor inputs_59_cast_fp16 = add(x = inputs_57_cast_fp16, y = obj_59_cast_fp16)[name = string("inputs_59_cast_fp16")]; + tensor out_59_axes_0 = const()[name = string("out_59_axes_0"), val = tensor([1])]; + fp16 var_1919_to_fp16 = const()[name = string("op_1919_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_59_cast_fp16 = layer_norm(axes = out_59_axes_0, epsilon = var_1919_to_fp16, x = inputs_59_cast_fp16)[name = string("out_59_cast_fp16")]; + tensor input_115_gamma_0_to_fp16 = const()[name = string("input_115_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(578730688)))]; + tensor input_115_beta_0_to_fp16 = const()[name = string("input_115_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(578733312)))]; + fp16 input_115_epsilon_0_to_fp16 = const()[name = string("input_115_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor 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 = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_59_cast_fp16)[name = string("input_115_cast_fp16")]; + string input_117_pad_type_0 = const()[name = string("input_117_pad_type_0"), val = string("valid")]; + tensor input_117_strides_0 = const()[name = string("input_117_strides_0"), val = tensor([1, 1])]; + tensor input_117_pad_0 = const()[name = string("input_117_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_117_dilations_0 = const()[name = string("input_117_dilations_0"), val = tensor([1, 1])]; + int32 input_117_groups_0 = const()[name = string("input_117_groups_0"), val = int32(1)]; + tensor layers_14_fc1_weight_to_fp16 = const()[name = string("layers_14_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(578735936)))]; + tensor layers_14_fc1_bias_to_fp16 = const()[name = string("layers_14_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(591843200)))]; + tensor input_117_cast_fp16 = conv(bias = layers_14_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_14_fc1_weight_to_fp16, x = input_115_cast_fp16)[name = string("input_117_cast_fp16")]; + string input_119_mode_0 = const()[name = string("input_119_mode_0"), val = string("EXACT")]; + tensor input_119_cast_fp16 = gelu(mode = input_119_mode_0, x = input_117_cast_fp16)[name = string("input_119_cast_fp16")]; + string hidden_states_33_pad_type_0 = const()[name = string("hidden_states_33_pad_type_0"), val = string("valid")]; + tensor hidden_states_33_strides_0 = const()[name = string("hidden_states_33_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_33_pad_0 = const()[name = string("hidden_states_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_33_dilations_0 = const()[name = string("hidden_states_33_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_33_groups_0 = const()[name = string("hidden_states_33_groups_0"), val = int32(1)]; + tensor layers_14_fc2_weight_to_fp16 = const()[name = string("layers_14_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(591853504)))]; + tensor layers_14_fc2_bias_to_fp16 = const()[name = string("layers_14_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604960768)))]; + tensor hidden_states_33_cast_fp16 = conv(bias = layers_14_fc2_bias_to_fp16, dilations = hidden_states_33_dilations_0, groups = hidden_states_33_groups_0, pad = hidden_states_33_pad_0, pad_type = hidden_states_33_pad_type_0, strides = hidden_states_33_strides_0, weight = layers_14_fc2_weight_to_fp16, x = input_119_cast_fp16)[name = string("hidden_states_33_cast_fp16")]; + tensor inputs_61_cast_fp16 = add(x = inputs_59_cast_fp16, y = hidden_states_33_cast_fp16)[name = string("inputs_61_cast_fp16")]; + int32 var_1952 = const()[name = string("op_1952"), val = int32(3)]; + tensor out_61_axes_0 = const()[name = string("out_61_axes_0"), val = tensor([1])]; + fp16 var_1971_to_fp16 = const()[name = string("op_1971_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_61_cast_fp16 = layer_norm(axes = out_61_axes_0, epsilon = var_1971_to_fp16, x = inputs_61_cast_fp16)[name = string("out_61_cast_fp16")]; + tensor obj_61_gamma_0_to_fp16 = const()[name = string("obj_61_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604963392)))]; + tensor obj_61_beta_0_to_fp16 = const()[name = string("obj_61_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604966016)))]; + fp16 obj_61_epsilon_0_to_fp16 = const()[name = string("obj_61_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor obj_61_cast_fp16 = batch_norm(beta = obj_61_beta_0_to_fp16, epsilon = obj_61_epsilon_0_to_fp16, gamma = obj_61_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_61_cast_fp16)[name = string("obj_61_cast_fp16")]; + string query_31_pad_type_0 = const()[name = string("query_31_pad_type_0"), val = string("valid")]; + tensor query_31_strides_0 = const()[name = string("query_31_strides_0"), val = tensor([1, 1])]; + tensor query_31_pad_0 = const()[name = string("query_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_31_dilations_0 = const()[name = string("query_31_dilations_0"), val = tensor([1, 1])]; + int32 query_31_groups_0 = const()[name = string("query_31_groups_0"), val = int32(1)]; + tensor layers_15_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_15_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604968640)))]; + tensor layers_15_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_15_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(608245504)))]; + tensor query_31_cast_fp16 = conv(bias = layers_15_self_attn_q_proj_bias_to_fp16, dilations = query_31_dilations_0, groups = query_31_groups_0, pad = query_31_pad_0, pad_type = query_31_pad_type_0, strides = query_31_strides_0, weight = layers_15_self_attn_q_proj_weight_to_fp16, x = obj_61_cast_fp16)[name = string("query_31_cast_fp16")]; + string key_31_pad_type_0 = const()[name = string("key_31_pad_type_0"), val = string("valid")]; + tensor key_31_strides_0 = const()[name = string("key_31_strides_0"), val = tensor([1, 1])]; + tensor key_31_pad_0 = const()[name = string("key_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_31_dilations_0 = const()[name = string("key_31_dilations_0"), val = tensor([1, 1])]; + int32 key_31_groups_0 = const()[name = string("key_31_groups_0"), val = int32(1)]; + tensor layers_15_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_15_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(608248128)))]; + tensor key_31_cast_fp16 = conv(dilations = key_31_dilations_0, groups = key_31_groups_0, pad = key_31_pad_0, pad_type = key_31_pad_type_0, strides = key_31_strides_0, weight = layers_15_self_attn_k_proj_weight_to_fp16, x = obj_61_cast_fp16)[name = string("key_31_cast_fp16")]; + string value_31_pad_type_0 = const()[name = string("value_31_pad_type_0"), val = string("valid")]; + tensor value_31_strides_0 = const()[name = string("value_31_strides_0"), val = tensor([1, 1])]; + tensor value_31_pad_0 = const()[name = string("value_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_31_dilations_0 = const()[name = string("value_31_dilations_0"), val = tensor([1, 1])]; + int32 value_31_groups_0 = const()[name = string("value_31_groups_0"), val = int32(1)]; + tensor layers_15_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_15_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(611524992)))]; + tensor layers_15_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_15_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(614801856)))]; + tensor value_31_cast_fp16 = conv(bias = layers_15_self_attn_v_proj_bias_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_61_cast_fp16)[name = string("value_31_cast_fp16")]; + tensor var_2006 = const()[name = string("op_2006"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_31_cast_fp16 = reshape(shape = var_2006, x = query_31_cast_fp16)[name = string("mh_q_31_cast_fp16")]; + fp16 var_2008_to_fp16 = const()[name = string("op_2008_to_fp16"), val = fp16(0x1p-3)]; + tensor var_2009_cast_fp16 = mul(x = mh_q_31_cast_fp16, y = var_2008_to_fp16)[name = string("op_2009_cast_fp16")]; + tensor var_2010 = const()[name = string("op_2010"), val = tensor([1, 20, 64, -1])]; + tensor var_2011_cast_fp16 = reshape(shape = var_2010, x = key_31_cast_fp16)[name = string("op_2011_cast_fp16")]; + bool mh_w_31_transpose_x_0 = const()[name = string("mh_w_31_transpose_x_0"), val = bool(true)]; + bool mh_w_31_transpose_y_0 = const()[name = string("mh_w_31_transpose_y_0"), val = bool(false)]; + tensor mh_w_31_cast_fp16 = matmul(transpose_x = mh_w_31_transpose_x_0, transpose_y = mh_w_31_transpose_y_0, x = var_2009_cast_fp16, y = var_2011_cast_fp16)[name = string("mh_w_31_cast_fp16")]; + tensor var_2014_cast_fp16 = softmax(axis = var_1952, x = mh_w_31_cast_fp16)[name = string("op_2014_cast_fp16")]; + tensor var_2015 = const()[name = string("op_2015"), val = tensor([1, 20, 64, -1])]; + tensor var_2016_cast_fp16 = reshape(shape = var_2015, x = value_31_cast_fp16)[name = string("op_2016_cast_fp16")]; + bool attn_31_transpose_x_0 = const()[name = string("attn_31_transpose_x_0"), val = bool(false)]; + bool attn_31_transpose_y_0 = const()[name = string("attn_31_transpose_y_0"), val = bool(true)]; + tensor attn_31_cast_fp16 = matmul(transpose_x = attn_31_transpose_x_0, transpose_y = attn_31_transpose_y_0, x = var_2016_cast_fp16, y = var_2014_cast_fp16)[name = string("attn_31_cast_fp16")]; + tensor var_2019 = const()[name = string("op_2019"), val = tensor([1, 1280, 1, -1])]; + tensor input_121_cast_fp16 = reshape(shape = var_2019, x = attn_31_cast_fp16)[name = string("input_121_cast_fp16")]; + string obj_63_pad_type_0 = const()[name = string("obj_63_pad_type_0"), val = string("valid")]; + tensor obj_63_strides_0 = const()[name = string("obj_63_strides_0"), val = tensor([1, 1])]; + tensor obj_63_pad_0 = const()[name = string("obj_63_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_63_dilations_0 = const()[name = string("obj_63_dilations_0"), val = tensor([1, 1])]; + int32 obj_63_groups_0 = const()[name = string("obj_63_groups_0"), val = int32(1)]; + tensor layers_15_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_15_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(614804480)))]; + tensor layers_15_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_15_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(618081344)))]; + tensor obj_63_cast_fp16 = conv(bias = layers_15_self_attn_o_proj_bias_to_fp16, dilations = obj_63_dilations_0, groups = obj_63_groups_0, pad = obj_63_pad_0, pad_type = obj_63_pad_type_0, strides = obj_63_strides_0, weight = layers_15_self_attn_o_proj_weight_to_fp16, x = input_121_cast_fp16)[name = string("obj_63_cast_fp16")]; + tensor inputs_63_cast_fp16 = add(x = inputs_61_cast_fp16, y = obj_63_cast_fp16)[name = string("inputs_63_cast_fp16")]; + tensor out_63_axes_0 = const()[name = string("out_63_axes_0"), val = tensor([1])]; + fp16 var_2037_to_fp16 = const()[name = string("op_2037_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_63_cast_fp16 = layer_norm(axes = out_63_axes_0, epsilon = var_2037_to_fp16, x = inputs_63_cast_fp16)[name = string("out_63_cast_fp16")]; + tensor input_123_gamma_0_to_fp16 = const()[name = string("input_123_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(618083968)))]; + tensor input_123_beta_0_to_fp16 = const()[name = string("input_123_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(618086592)))]; + fp16 input_123_epsilon_0_to_fp16 = const()[name = string("input_123_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor input_123_cast_fp16 = batch_norm(beta = input_123_beta_0_to_fp16, epsilon = input_123_epsilon_0_to_fp16, gamma = input_123_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_63_cast_fp16)[name = string("input_123_cast_fp16")]; + string input_125_pad_type_0 = const()[name = string("input_125_pad_type_0"), val = string("valid")]; + tensor input_125_strides_0 = const()[name = string("input_125_strides_0"), val = tensor([1, 1])]; + tensor input_125_pad_0 = const()[name = string("input_125_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_125_dilations_0 = const()[name = string("input_125_dilations_0"), val = tensor([1, 1])]; + int32 input_125_groups_0 = const()[name = string("input_125_groups_0"), val = int32(1)]; + tensor layers_15_fc1_weight_to_fp16 = const()[name = string("layers_15_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(618089216)))]; + tensor layers_15_fc1_bias_to_fp16 = const()[name = string("layers_15_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631196480)))]; + tensor input_125_cast_fp16 = conv(bias = layers_15_fc1_bias_to_fp16, dilations = input_125_dilations_0, groups = input_125_groups_0, pad = input_125_pad_0, pad_type = input_125_pad_type_0, strides = input_125_strides_0, weight = layers_15_fc1_weight_to_fp16, x = input_123_cast_fp16)[name = string("input_125_cast_fp16")]; + string input_127_mode_0 = const()[name = string("input_127_mode_0"), val = string("EXACT")]; + tensor input_127_cast_fp16 = gelu(mode = input_127_mode_0, x = input_125_cast_fp16)[name = string("input_127_cast_fp16")]; + string hidden_states_35_pad_type_0 = const()[name = string("hidden_states_35_pad_type_0"), val = string("valid")]; + tensor hidden_states_35_strides_0 = const()[name = string("hidden_states_35_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_35_pad_0 = const()[name = string("hidden_states_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_35_dilations_0 = const()[name = string("hidden_states_35_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_35_groups_0 = const()[name = string("hidden_states_35_groups_0"), val = int32(1)]; + tensor layers_15_fc2_weight_to_fp16 = const()[name = string("layers_15_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631206784)))]; + tensor layers_15_fc2_bias_to_fp16 = const()[name = string("layers_15_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(644314048)))]; + tensor hidden_states_35_cast_fp16 = conv(bias = layers_15_fc2_bias_to_fp16, dilations = hidden_states_35_dilations_0, groups = hidden_states_35_groups_0, pad = hidden_states_35_pad_0, pad_type = hidden_states_35_pad_type_0, strides = hidden_states_35_strides_0, weight = layers_15_fc2_weight_to_fp16, x = input_127_cast_fp16)[name = string("hidden_states_35_cast_fp16")]; + tensor inputs_65_cast_fp16 = add(x = inputs_63_cast_fp16, y = hidden_states_35_cast_fp16)[name = string("inputs_65_cast_fp16")]; + int32 var_2070 = const()[name = string("op_2070"), val = int32(3)]; + tensor out_65_axes_0 = const()[name = string("out_65_axes_0"), val = tensor([1])]; + fp16 var_2089_to_fp16 = const()[name = string("op_2089_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_65_cast_fp16 = layer_norm(axes = out_65_axes_0, epsilon = var_2089_to_fp16, x = inputs_65_cast_fp16)[name = string("out_65_cast_fp16")]; + tensor obj_65_gamma_0_to_fp16 = const()[name = string("obj_65_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(644316672)))]; + tensor obj_65_beta_0_to_fp16 = const()[name = string("obj_65_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(644319296)))]; + fp16 obj_65_epsilon_0_to_fp16 = const()[name = string("obj_65_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor obj_65_cast_fp16 = batch_norm(beta = obj_65_beta_0_to_fp16, epsilon = obj_65_epsilon_0_to_fp16, gamma = obj_65_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_65_cast_fp16)[name = string("obj_65_cast_fp16")]; + string query_33_pad_type_0 = const()[name = string("query_33_pad_type_0"), val = string("valid")]; + tensor query_33_strides_0 = const()[name = string("query_33_strides_0"), val = tensor([1, 1])]; + tensor query_33_pad_0 = const()[name = string("query_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_33_dilations_0 = const()[name = string("query_33_dilations_0"), val = tensor([1, 1])]; + int32 query_33_groups_0 = const()[name = string("query_33_groups_0"), val = int32(1)]; + tensor layers_16_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_16_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(644321920)))]; + tensor layers_16_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_16_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(647598784)))]; + tensor query_33_cast_fp16 = conv(bias = layers_16_self_attn_q_proj_bias_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_16_self_attn_q_proj_weight_to_fp16, x = obj_65_cast_fp16)[name = string("query_33_cast_fp16")]; + string key_33_pad_type_0 = const()[name = string("key_33_pad_type_0"), val = string("valid")]; + tensor key_33_strides_0 = const()[name = string("key_33_strides_0"), val = tensor([1, 1])]; + tensor key_33_pad_0 = const()[name = string("key_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_33_dilations_0 = const()[name = string("key_33_dilations_0"), val = tensor([1, 1])]; + int32 key_33_groups_0 = const()[name = string("key_33_groups_0"), val = int32(1)]; + tensor layers_16_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_16_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(647601408)))]; + tensor key_33_cast_fp16 = conv(dilations = key_33_dilations_0, groups = key_33_groups_0, pad = key_33_pad_0, pad_type = key_33_pad_type_0, strides = key_33_strides_0, weight = layers_16_self_attn_k_proj_weight_to_fp16, x = obj_65_cast_fp16)[name = string("key_33_cast_fp16")]; + string value_33_pad_type_0 = const()[name = string("value_33_pad_type_0"), val = string("valid")]; + tensor value_33_strides_0 = const()[name = string("value_33_strides_0"), val = tensor([1, 1])]; + tensor value_33_pad_0 = const()[name = string("value_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_33_dilations_0 = const()[name = string("value_33_dilations_0"), val = tensor([1, 1])]; + int32 value_33_groups_0 = const()[name = string("value_33_groups_0"), val = int32(1)]; + tensor layers_16_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_16_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(650878272)))]; + tensor layers_16_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_16_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(654155136)))]; + tensor value_33_cast_fp16 = conv(bias = layers_16_self_attn_v_proj_bias_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_65_cast_fp16)[name = string("value_33_cast_fp16")]; + tensor var_2124 = const()[name = string("op_2124"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_33_cast_fp16 = reshape(shape = var_2124, x = query_33_cast_fp16)[name = string("mh_q_33_cast_fp16")]; + fp16 var_2126_to_fp16 = const()[name = string("op_2126_to_fp16"), val = fp16(0x1p-3)]; + tensor var_2127_cast_fp16 = mul(x = mh_q_33_cast_fp16, y = var_2126_to_fp16)[name = string("op_2127_cast_fp16")]; + tensor var_2128 = const()[name = string("op_2128"), val = tensor([1, 20, 64, -1])]; + tensor var_2129_cast_fp16 = reshape(shape = var_2128, x = key_33_cast_fp16)[name = string("op_2129_cast_fp16")]; + bool mh_w_33_transpose_x_0 = const()[name = string("mh_w_33_transpose_x_0"), val = bool(true)]; + bool mh_w_33_transpose_y_0 = const()[name = string("mh_w_33_transpose_y_0"), val = bool(false)]; + tensor mh_w_33_cast_fp16 = matmul(transpose_x = mh_w_33_transpose_x_0, transpose_y = mh_w_33_transpose_y_0, x = var_2127_cast_fp16, y = var_2129_cast_fp16)[name = string("mh_w_33_cast_fp16")]; + tensor var_2132_cast_fp16 = softmax(axis = var_2070, x = mh_w_33_cast_fp16)[name = string("op_2132_cast_fp16")]; + tensor var_2133 = const()[name = string("op_2133"), val = tensor([1, 20, 64, -1])]; + tensor var_2134_cast_fp16 = reshape(shape = var_2133, x = value_33_cast_fp16)[name = string("op_2134_cast_fp16")]; + bool attn_33_transpose_x_0 = const()[name = string("attn_33_transpose_x_0"), val = bool(false)]; + bool attn_33_transpose_y_0 = const()[name = string("attn_33_transpose_y_0"), val = bool(true)]; + tensor attn_33_cast_fp16 = matmul(transpose_x = attn_33_transpose_x_0, transpose_y = attn_33_transpose_y_0, x = var_2134_cast_fp16, y = var_2132_cast_fp16)[name = string("attn_33_cast_fp16")]; + tensor var_2137 = const()[name = string("op_2137"), val = tensor([1, 1280, 1, -1])]; + tensor input_129_cast_fp16 = reshape(shape = var_2137, x = attn_33_cast_fp16)[name = string("input_129_cast_fp16")]; + string obj_67_pad_type_0 = const()[name = string("obj_67_pad_type_0"), val = string("valid")]; + tensor obj_67_strides_0 = const()[name = string("obj_67_strides_0"), val = tensor([1, 1])]; + tensor obj_67_pad_0 = const()[name = string("obj_67_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_67_dilations_0 = const()[name = string("obj_67_dilations_0"), val = tensor([1, 1])]; + int32 obj_67_groups_0 = const()[name = string("obj_67_groups_0"), val = int32(1)]; + tensor layers_16_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_16_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(654157760)))]; + tensor layers_16_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_16_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(657434624)))]; + tensor obj_67_cast_fp16 = conv(bias = layers_16_self_attn_o_proj_bias_to_fp16, dilations = obj_67_dilations_0, groups = obj_67_groups_0, pad = obj_67_pad_0, pad_type = obj_67_pad_type_0, strides = obj_67_strides_0, weight = layers_16_self_attn_o_proj_weight_to_fp16, x = input_129_cast_fp16)[name = string("obj_67_cast_fp16")]; + tensor inputs_67_cast_fp16 = add(x = inputs_65_cast_fp16, y = obj_67_cast_fp16)[name = string("inputs_67_cast_fp16")]; + tensor out_67_axes_0 = const()[name = string("out_67_axes_0"), val = tensor([1])]; + fp16 var_2155_to_fp16 = const()[name = string("op_2155_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_67_cast_fp16 = layer_norm(axes = out_67_axes_0, epsilon = var_2155_to_fp16, x = inputs_67_cast_fp16)[name = string("out_67_cast_fp16")]; + tensor input_131_gamma_0_to_fp16 = const()[name = string("input_131_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(657437248)))]; + tensor input_131_beta_0_to_fp16 = const()[name = string("input_131_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(657439872)))]; + fp16 input_131_epsilon_0_to_fp16 = const()[name = string("input_131_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor input_131_cast_fp16 = batch_norm(beta = input_131_beta_0_to_fp16, epsilon = input_131_epsilon_0_to_fp16, gamma = input_131_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_67_cast_fp16)[name = string("input_131_cast_fp16")]; + string input_133_pad_type_0 = const()[name = string("input_133_pad_type_0"), val = string("valid")]; + tensor input_133_strides_0 = const()[name = string("input_133_strides_0"), val = tensor([1, 1])]; + tensor input_133_pad_0 = const()[name = string("input_133_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_133_dilations_0 = const()[name = string("input_133_dilations_0"), val = tensor([1, 1])]; + int32 input_133_groups_0 = const()[name = string("input_133_groups_0"), val = int32(1)]; + tensor layers_16_fc1_weight_to_fp16 = const()[name = string("layers_16_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(657442496)))]; + tensor layers_16_fc1_bias_to_fp16 = const()[name = string("layers_16_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(670549760)))]; + tensor input_133_cast_fp16 = conv(bias = layers_16_fc1_bias_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 = layers_16_fc1_weight_to_fp16, x = input_131_cast_fp16)[name = string("input_133_cast_fp16")]; + string input_135_mode_0 = const()[name = string("input_135_mode_0"), val = string("EXACT")]; + tensor input_135_cast_fp16 = gelu(mode = input_135_mode_0, x = input_133_cast_fp16)[name = string("input_135_cast_fp16")]; + string hidden_states_37_pad_type_0 = const()[name = string("hidden_states_37_pad_type_0"), val = string("valid")]; + tensor hidden_states_37_strides_0 = const()[name = string("hidden_states_37_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_37_pad_0 = const()[name = string("hidden_states_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_37_dilations_0 = const()[name = string("hidden_states_37_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_37_groups_0 = const()[name = string("hidden_states_37_groups_0"), val = int32(1)]; + tensor layers_16_fc2_weight_to_fp16 = const()[name = string("layers_16_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(670560064)))]; + tensor layers_16_fc2_bias_to_fp16 = const()[name = string("layers_16_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(683667328)))]; + tensor hidden_states_37_cast_fp16 = conv(bias = layers_16_fc2_bias_to_fp16, dilations = hidden_states_37_dilations_0, groups = hidden_states_37_groups_0, pad = hidden_states_37_pad_0, pad_type = hidden_states_37_pad_type_0, strides = hidden_states_37_strides_0, weight = layers_16_fc2_weight_to_fp16, x = input_135_cast_fp16)[name = string("hidden_states_37_cast_fp16")]; + tensor inputs_69_cast_fp16 = add(x = inputs_67_cast_fp16, y = hidden_states_37_cast_fp16)[name = string("inputs_69_cast_fp16")]; + int32 var_2188 = const()[name = string("op_2188"), val = int32(3)]; + tensor out_69_axes_0 = const()[name = string("out_69_axes_0"), val = tensor([1])]; + fp16 var_2207_to_fp16 = const()[name = string("op_2207_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_69_cast_fp16 = layer_norm(axes = out_69_axes_0, epsilon = var_2207_to_fp16, x = inputs_69_cast_fp16)[name = string("out_69_cast_fp16")]; + tensor obj_69_gamma_0_to_fp16 = const()[name = string("obj_69_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(683669952)))]; + tensor obj_69_beta_0_to_fp16 = const()[name = string("obj_69_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(683672576)))]; + fp16 obj_69_epsilon_0_to_fp16 = const()[name = string("obj_69_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor obj_69_cast_fp16 = batch_norm(beta = obj_69_beta_0_to_fp16, epsilon = obj_69_epsilon_0_to_fp16, gamma = obj_69_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_69_cast_fp16)[name = string("obj_69_cast_fp16")]; + string query_35_pad_type_0 = const()[name = string("query_35_pad_type_0"), val = string("valid")]; + tensor query_35_strides_0 = const()[name = string("query_35_strides_0"), val = tensor([1, 1])]; + tensor query_35_pad_0 = const()[name = string("query_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_35_dilations_0 = const()[name = string("query_35_dilations_0"), val = tensor([1, 1])]; + int32 query_35_groups_0 = const()[name = string("query_35_groups_0"), val = int32(1)]; + tensor layers_17_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_17_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(683675200)))]; + tensor layers_17_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_17_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686952064)))]; + tensor query_35_cast_fp16 = conv(bias = layers_17_self_attn_q_proj_bias_to_fp16, dilations = query_35_dilations_0, groups = query_35_groups_0, pad = query_35_pad_0, pad_type = query_35_pad_type_0, strides = query_35_strides_0, weight = layers_17_self_attn_q_proj_weight_to_fp16, x = obj_69_cast_fp16)[name = string("query_35_cast_fp16")]; + string key_35_pad_type_0 = const()[name = string("key_35_pad_type_0"), val = string("valid")]; + tensor key_35_strides_0 = const()[name = string("key_35_strides_0"), val = tensor([1, 1])]; + tensor key_35_pad_0 = const()[name = string("key_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_35_dilations_0 = const()[name = string("key_35_dilations_0"), val = tensor([1, 1])]; + int32 key_35_groups_0 = const()[name = string("key_35_groups_0"), val = int32(1)]; + tensor layers_17_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_17_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686954688)))]; + tensor key_35_cast_fp16 = conv(dilations = key_35_dilations_0, groups = key_35_groups_0, pad = key_35_pad_0, pad_type = key_35_pad_type_0, strides = key_35_strides_0, weight = layers_17_self_attn_k_proj_weight_to_fp16, x = obj_69_cast_fp16)[name = string("key_35_cast_fp16")]; + string value_35_pad_type_0 = const()[name = string("value_35_pad_type_0"), val = string("valid")]; + tensor value_35_strides_0 = const()[name = string("value_35_strides_0"), val = tensor([1, 1])]; + tensor value_35_pad_0 = const()[name = string("value_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_35_dilations_0 = const()[name = string("value_35_dilations_0"), val = tensor([1, 1])]; + int32 value_35_groups_0 = const()[name = string("value_35_groups_0"), val = int32(1)]; + tensor layers_17_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_17_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690231552)))]; + tensor layers_17_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_17_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(693508416)))]; + tensor value_35_cast_fp16 = conv(bias = layers_17_self_attn_v_proj_bias_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_69_cast_fp16)[name = string("value_35_cast_fp16")]; + tensor var_2242 = const()[name = string("op_2242"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_35_cast_fp16 = reshape(shape = var_2242, x = query_35_cast_fp16)[name = string("mh_q_35_cast_fp16")]; + fp16 var_2244_to_fp16 = const()[name = string("op_2244_to_fp16"), val = fp16(0x1p-3)]; + tensor var_2245_cast_fp16 = mul(x = mh_q_35_cast_fp16, y = var_2244_to_fp16)[name = string("op_2245_cast_fp16")]; + tensor var_2246 = const()[name = string("op_2246"), val = tensor([1, 20, 64, -1])]; + tensor var_2247_cast_fp16 = reshape(shape = var_2246, x = key_35_cast_fp16)[name = string("op_2247_cast_fp16")]; + bool mh_w_35_transpose_x_0 = const()[name = string("mh_w_35_transpose_x_0"), val = bool(true)]; + bool mh_w_35_transpose_y_0 = const()[name = string("mh_w_35_transpose_y_0"), val = bool(false)]; + tensor mh_w_35_cast_fp16 = matmul(transpose_x = mh_w_35_transpose_x_0, transpose_y = mh_w_35_transpose_y_0, x = var_2245_cast_fp16, y = var_2247_cast_fp16)[name = string("mh_w_35_cast_fp16")]; + tensor var_2250_cast_fp16 = softmax(axis = var_2188, x = mh_w_35_cast_fp16)[name = string("op_2250_cast_fp16")]; + tensor var_2251 = const()[name = string("op_2251"), val = tensor([1, 20, 64, -1])]; + tensor var_2252_cast_fp16 = reshape(shape = var_2251, x = value_35_cast_fp16)[name = string("op_2252_cast_fp16")]; + bool attn_35_transpose_x_0 = const()[name = string("attn_35_transpose_x_0"), val = bool(false)]; + bool attn_35_transpose_y_0 = const()[name = string("attn_35_transpose_y_0"), val = bool(true)]; + tensor attn_35_cast_fp16 = matmul(transpose_x = attn_35_transpose_x_0, transpose_y = attn_35_transpose_y_0, x = var_2252_cast_fp16, y = var_2250_cast_fp16)[name = string("attn_35_cast_fp16")]; + tensor var_2255 = const()[name = string("op_2255"), val = tensor([1, 1280, 1, -1])]; + tensor input_137_cast_fp16 = reshape(shape = var_2255, x = attn_35_cast_fp16)[name = string("input_137_cast_fp16")]; + string obj_71_pad_type_0 = const()[name = string("obj_71_pad_type_0"), val = string("valid")]; + tensor obj_71_strides_0 = const()[name = string("obj_71_strides_0"), val = tensor([1, 1])]; + tensor obj_71_pad_0 = const()[name = string("obj_71_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_71_dilations_0 = const()[name = string("obj_71_dilations_0"), val = tensor([1, 1])]; + int32 obj_71_groups_0 = const()[name = string("obj_71_groups_0"), val = int32(1)]; + tensor layers_17_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_17_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(693511040)))]; + tensor layers_17_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_17_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(696787904)))]; + tensor obj_71_cast_fp16 = conv(bias = layers_17_self_attn_o_proj_bias_to_fp16, dilations = obj_71_dilations_0, groups = obj_71_groups_0, pad = obj_71_pad_0, pad_type = obj_71_pad_type_0, strides = obj_71_strides_0, weight = layers_17_self_attn_o_proj_weight_to_fp16, x = input_137_cast_fp16)[name = string("obj_71_cast_fp16")]; + tensor inputs_71_cast_fp16 = add(x = inputs_69_cast_fp16, y = obj_71_cast_fp16)[name = string("inputs_71_cast_fp16")]; + tensor out_71_axes_0 = const()[name = string("out_71_axes_0"), val = tensor([1])]; + fp16 var_2273_to_fp16 = const()[name = string("op_2273_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_71_cast_fp16 = layer_norm(axes = out_71_axes_0, epsilon = var_2273_to_fp16, x = inputs_71_cast_fp16)[name = string("out_71_cast_fp16")]; + tensor input_139_gamma_0_to_fp16 = const()[name = string("input_139_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(696790528)))]; + tensor input_139_beta_0_to_fp16 = const()[name = string("input_139_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(696793152)))]; + fp16 input_139_epsilon_0_to_fp16 = const()[name = string("input_139_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor 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 = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_71_cast_fp16)[name = string("input_139_cast_fp16")]; + string input_141_pad_type_0 = const()[name = string("input_141_pad_type_0"), val = string("valid")]; + tensor input_141_strides_0 = const()[name = string("input_141_strides_0"), val = tensor([1, 1])]; + tensor input_141_pad_0 = const()[name = string("input_141_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_141_dilations_0 = const()[name = string("input_141_dilations_0"), val = tensor([1, 1])]; + int32 input_141_groups_0 = const()[name = string("input_141_groups_0"), val = int32(1)]; + tensor layers_17_fc1_weight_to_fp16 = const()[name = string("layers_17_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(696795776)))]; + tensor layers_17_fc1_bias_to_fp16 = const()[name = string("layers_17_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(709903040)))]; + tensor input_141_cast_fp16 = conv(bias = layers_17_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_17_fc1_weight_to_fp16, x = input_139_cast_fp16)[name = string("input_141_cast_fp16")]; + string input_143_mode_0 = const()[name = string("input_143_mode_0"), val = string("EXACT")]; + tensor input_143_cast_fp16 = gelu(mode = input_143_mode_0, x = input_141_cast_fp16)[name = string("input_143_cast_fp16")]; + string hidden_states_39_pad_type_0 = const()[name = string("hidden_states_39_pad_type_0"), val = string("valid")]; + tensor hidden_states_39_strides_0 = const()[name = string("hidden_states_39_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_39_pad_0 = const()[name = string("hidden_states_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_39_dilations_0 = const()[name = string("hidden_states_39_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_39_groups_0 = const()[name = string("hidden_states_39_groups_0"), val = int32(1)]; + tensor layers_17_fc2_weight_to_fp16 = const()[name = string("layers_17_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(709913344)))]; + tensor layers_17_fc2_bias_to_fp16 = const()[name = string("layers_17_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(723020608)))]; + tensor hidden_states_39_cast_fp16 = conv(bias = layers_17_fc2_bias_to_fp16, dilations = hidden_states_39_dilations_0, groups = hidden_states_39_groups_0, pad = hidden_states_39_pad_0, pad_type = hidden_states_39_pad_type_0, strides = hidden_states_39_strides_0, weight = layers_17_fc2_weight_to_fp16, x = input_143_cast_fp16)[name = string("hidden_states_39_cast_fp16")]; + tensor inputs_73_cast_fp16 = add(x = inputs_71_cast_fp16, y = hidden_states_39_cast_fp16)[name = string("inputs_73_cast_fp16")]; + int32 var_2306 = const()[name = string("op_2306"), val = int32(3)]; + tensor out_73_axes_0 = const()[name = string("out_73_axes_0"), val = tensor([1])]; + fp16 var_2325_to_fp16 = const()[name = string("op_2325_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_73_cast_fp16 = layer_norm(axes = out_73_axes_0, epsilon = var_2325_to_fp16, x = inputs_73_cast_fp16)[name = string("out_73_cast_fp16")]; + tensor obj_73_gamma_0_to_fp16 = const()[name = string("obj_73_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(723023232)))]; + tensor obj_73_beta_0_to_fp16 = const()[name = string("obj_73_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(723025856)))]; + fp16 obj_73_epsilon_0_to_fp16 = const()[name = string("obj_73_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor obj_73_cast_fp16 = batch_norm(beta = obj_73_beta_0_to_fp16, epsilon = obj_73_epsilon_0_to_fp16, gamma = obj_73_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_73_cast_fp16)[name = string("obj_73_cast_fp16")]; + string query_37_pad_type_0 = const()[name = string("query_37_pad_type_0"), val = string("valid")]; + tensor query_37_strides_0 = const()[name = string("query_37_strides_0"), val = tensor([1, 1])]; + tensor query_37_pad_0 = const()[name = string("query_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_37_dilations_0 = const()[name = string("query_37_dilations_0"), val = tensor([1, 1])]; + int32 query_37_groups_0 = const()[name = string("query_37_groups_0"), val = int32(1)]; + tensor layers_18_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_18_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(723028480)))]; + tensor layers_18_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_18_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(726305344)))]; + tensor query_37_cast_fp16 = conv(bias = layers_18_self_attn_q_proj_bias_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_18_self_attn_q_proj_weight_to_fp16, x = obj_73_cast_fp16)[name = string("query_37_cast_fp16")]; + string key_37_pad_type_0 = const()[name = string("key_37_pad_type_0"), val = string("valid")]; + tensor key_37_strides_0 = const()[name = string("key_37_strides_0"), val = tensor([1, 1])]; + tensor key_37_pad_0 = const()[name = string("key_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_37_dilations_0 = const()[name = string("key_37_dilations_0"), val = tensor([1, 1])]; + int32 key_37_groups_0 = const()[name = string("key_37_groups_0"), val = int32(1)]; + tensor layers_18_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_18_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(726307968)))]; + tensor key_37_cast_fp16 = conv(dilations = key_37_dilations_0, groups = key_37_groups_0, pad = key_37_pad_0, pad_type = key_37_pad_type_0, strides = key_37_strides_0, weight = layers_18_self_attn_k_proj_weight_to_fp16, x = obj_73_cast_fp16)[name = string("key_37_cast_fp16")]; + string value_37_pad_type_0 = const()[name = string("value_37_pad_type_0"), val = string("valid")]; + tensor value_37_strides_0 = const()[name = string("value_37_strides_0"), val = tensor([1, 1])]; + tensor value_37_pad_0 = const()[name = string("value_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_37_dilations_0 = const()[name = string("value_37_dilations_0"), val = tensor([1, 1])]; + int32 value_37_groups_0 = const()[name = string("value_37_groups_0"), val = int32(1)]; + tensor layers_18_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_18_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(729584832)))]; + tensor layers_18_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_18_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(732861696)))]; + tensor value_37_cast_fp16 = conv(bias = layers_18_self_attn_v_proj_bias_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_73_cast_fp16)[name = string("value_37_cast_fp16")]; + tensor var_2360 = const()[name = string("op_2360"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_37_cast_fp16 = reshape(shape = var_2360, x = query_37_cast_fp16)[name = string("mh_q_37_cast_fp16")]; + fp16 var_2362_to_fp16 = const()[name = string("op_2362_to_fp16"), val = fp16(0x1p-3)]; + tensor var_2363_cast_fp16 = mul(x = mh_q_37_cast_fp16, y = var_2362_to_fp16)[name = string("op_2363_cast_fp16")]; + tensor var_2364 = const()[name = string("op_2364"), val = tensor([1, 20, 64, -1])]; + tensor var_2365_cast_fp16 = reshape(shape = var_2364, x = key_37_cast_fp16)[name = string("op_2365_cast_fp16")]; + bool mh_w_37_transpose_x_0 = const()[name = string("mh_w_37_transpose_x_0"), val = bool(true)]; + bool mh_w_37_transpose_y_0 = const()[name = string("mh_w_37_transpose_y_0"), val = bool(false)]; + tensor mh_w_37_cast_fp16 = matmul(transpose_x = mh_w_37_transpose_x_0, transpose_y = mh_w_37_transpose_y_0, x = var_2363_cast_fp16, y = var_2365_cast_fp16)[name = string("mh_w_37_cast_fp16")]; + tensor var_2368_cast_fp16 = softmax(axis = var_2306, x = mh_w_37_cast_fp16)[name = string("op_2368_cast_fp16")]; + tensor var_2369 = const()[name = string("op_2369"), val = tensor([1, 20, 64, -1])]; + tensor var_2370_cast_fp16 = reshape(shape = var_2369, x = value_37_cast_fp16)[name = string("op_2370_cast_fp16")]; + bool attn_37_transpose_x_0 = const()[name = string("attn_37_transpose_x_0"), val = bool(false)]; + bool attn_37_transpose_y_0 = const()[name = string("attn_37_transpose_y_0"), val = bool(true)]; + tensor attn_37_cast_fp16 = matmul(transpose_x = attn_37_transpose_x_0, transpose_y = attn_37_transpose_y_0, x = var_2370_cast_fp16, y = var_2368_cast_fp16)[name = string("attn_37_cast_fp16")]; + tensor var_2373 = const()[name = string("op_2373"), val = tensor([1, 1280, 1, -1])]; + tensor input_145_cast_fp16 = reshape(shape = var_2373, x = attn_37_cast_fp16)[name = string("input_145_cast_fp16")]; + string obj_75_pad_type_0 = const()[name = string("obj_75_pad_type_0"), val = string("valid")]; + tensor obj_75_strides_0 = const()[name = string("obj_75_strides_0"), val = tensor([1, 1])]; + tensor obj_75_pad_0 = const()[name = string("obj_75_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_75_dilations_0 = const()[name = string("obj_75_dilations_0"), val = tensor([1, 1])]; + int32 obj_75_groups_0 = const()[name = string("obj_75_groups_0"), val = int32(1)]; + tensor layers_18_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_18_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(732864320)))]; + tensor layers_18_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_18_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(736141184)))]; + tensor obj_75_cast_fp16 = conv(bias = layers_18_self_attn_o_proj_bias_to_fp16, dilations = obj_75_dilations_0, groups = obj_75_groups_0, pad = obj_75_pad_0, pad_type = obj_75_pad_type_0, strides = obj_75_strides_0, weight = layers_18_self_attn_o_proj_weight_to_fp16, x = input_145_cast_fp16)[name = string("obj_75_cast_fp16")]; + tensor inputs_75_cast_fp16 = add(x = inputs_73_cast_fp16, y = obj_75_cast_fp16)[name = string("inputs_75_cast_fp16")]; + tensor out_75_axes_0 = const()[name = string("out_75_axes_0"), val = tensor([1])]; + fp16 var_2391_to_fp16 = const()[name = string("op_2391_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_75_cast_fp16 = layer_norm(axes = out_75_axes_0, epsilon = var_2391_to_fp16, x = inputs_75_cast_fp16)[name = string("out_75_cast_fp16")]; + tensor input_147_gamma_0_to_fp16 = const()[name = string("input_147_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(736143808)))]; + tensor input_147_beta_0_to_fp16 = const()[name = string("input_147_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(736146432)))]; + fp16 input_147_epsilon_0_to_fp16 = const()[name = string("input_147_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor 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 = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_75_cast_fp16)[name = string("input_147_cast_fp16")]; + string input_149_pad_type_0 = const()[name = string("input_149_pad_type_0"), val = string("valid")]; + tensor input_149_strides_0 = const()[name = string("input_149_strides_0"), val = tensor([1, 1])]; + tensor input_149_pad_0 = const()[name = string("input_149_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_149_dilations_0 = const()[name = string("input_149_dilations_0"), val = tensor([1, 1])]; + int32 input_149_groups_0 = const()[name = string("input_149_groups_0"), val = int32(1)]; + tensor layers_18_fc1_weight_to_fp16 = const()[name = string("layers_18_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(736149056)))]; + tensor layers_18_fc1_bias_to_fp16 = const()[name = string("layers_18_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(749256320)))]; + tensor input_149_cast_fp16 = conv(bias = layers_18_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_18_fc1_weight_to_fp16, x = input_147_cast_fp16)[name = string("input_149_cast_fp16")]; + string input_151_mode_0 = const()[name = string("input_151_mode_0"), val = string("EXACT")]; + tensor input_151_cast_fp16 = gelu(mode = input_151_mode_0, x = input_149_cast_fp16)[name = string("input_151_cast_fp16")]; + string hidden_states_41_pad_type_0 = const()[name = string("hidden_states_41_pad_type_0"), val = string("valid")]; + tensor hidden_states_41_strides_0 = const()[name = string("hidden_states_41_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_41_pad_0 = const()[name = string("hidden_states_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_41_dilations_0 = const()[name = string("hidden_states_41_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_41_groups_0 = const()[name = string("hidden_states_41_groups_0"), val = int32(1)]; + tensor layers_18_fc2_weight_to_fp16 = const()[name = string("layers_18_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(749266624)))]; + tensor layers_18_fc2_bias_to_fp16 = const()[name = string("layers_18_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(762373888)))]; + tensor hidden_states_41_cast_fp16 = conv(bias = layers_18_fc2_bias_to_fp16, dilations = hidden_states_41_dilations_0, groups = hidden_states_41_groups_0, pad = hidden_states_41_pad_0, pad_type = hidden_states_41_pad_type_0, strides = hidden_states_41_strides_0, weight = layers_18_fc2_weight_to_fp16, x = input_151_cast_fp16)[name = string("hidden_states_41_cast_fp16")]; + tensor inputs_77_cast_fp16 = add(x = inputs_75_cast_fp16, y = hidden_states_41_cast_fp16)[name = string("inputs_77_cast_fp16")]; + int32 var_2424 = const()[name = string("op_2424"), val = int32(3)]; + tensor out_77_axes_0 = const()[name = string("out_77_axes_0"), val = tensor([1])]; + fp16 var_2443_to_fp16 = const()[name = string("op_2443_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_77_cast_fp16 = layer_norm(axes = out_77_axes_0, epsilon = var_2443_to_fp16, x = inputs_77_cast_fp16)[name = string("out_77_cast_fp16")]; + tensor obj_77_gamma_0_to_fp16 = const()[name = string("obj_77_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(762376512)))]; + tensor obj_77_beta_0_to_fp16 = const()[name = string("obj_77_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(762379136)))]; + fp16 obj_77_epsilon_0_to_fp16 = const()[name = string("obj_77_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor obj_77_cast_fp16 = batch_norm(beta = obj_77_beta_0_to_fp16, epsilon = obj_77_epsilon_0_to_fp16, gamma = obj_77_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_77_cast_fp16)[name = string("obj_77_cast_fp16")]; + string query_39_pad_type_0 = const()[name = string("query_39_pad_type_0"), val = string("valid")]; + tensor query_39_strides_0 = const()[name = string("query_39_strides_0"), val = tensor([1, 1])]; + tensor query_39_pad_0 = const()[name = string("query_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_39_dilations_0 = const()[name = string("query_39_dilations_0"), val = tensor([1, 1])]; + int32 query_39_groups_0 = const()[name = string("query_39_groups_0"), val = int32(1)]; + tensor layers_19_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_19_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(762381760)))]; + tensor layers_19_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_19_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(765658624)))]; + tensor query_39_cast_fp16 = conv(bias = layers_19_self_attn_q_proj_bias_to_fp16, dilations = query_39_dilations_0, groups = query_39_groups_0, pad = query_39_pad_0, pad_type = query_39_pad_type_0, strides = query_39_strides_0, weight = layers_19_self_attn_q_proj_weight_to_fp16, x = obj_77_cast_fp16)[name = string("query_39_cast_fp16")]; + string key_39_pad_type_0 = const()[name = string("key_39_pad_type_0"), val = string("valid")]; + tensor key_39_strides_0 = const()[name = string("key_39_strides_0"), val = tensor([1, 1])]; + tensor key_39_pad_0 = const()[name = string("key_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_39_dilations_0 = const()[name = string("key_39_dilations_0"), val = tensor([1, 1])]; + int32 key_39_groups_0 = const()[name = string("key_39_groups_0"), val = int32(1)]; + tensor layers_19_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_19_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(765661248)))]; + tensor key_39_cast_fp16 = conv(dilations = key_39_dilations_0, groups = key_39_groups_0, pad = key_39_pad_0, pad_type = key_39_pad_type_0, strides = key_39_strides_0, weight = layers_19_self_attn_k_proj_weight_to_fp16, x = obj_77_cast_fp16)[name = string("key_39_cast_fp16")]; + string value_39_pad_type_0 = const()[name = string("value_39_pad_type_0"), val = string("valid")]; + tensor value_39_strides_0 = const()[name = string("value_39_strides_0"), val = tensor([1, 1])]; + tensor value_39_pad_0 = const()[name = string("value_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_39_dilations_0 = const()[name = string("value_39_dilations_0"), val = tensor([1, 1])]; + int32 value_39_groups_0 = const()[name = string("value_39_groups_0"), val = int32(1)]; + tensor layers_19_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_19_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(768938112)))]; + tensor layers_19_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_19_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(772214976)))]; + tensor value_39_cast_fp16 = conv(bias = layers_19_self_attn_v_proj_bias_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_77_cast_fp16)[name = string("value_39_cast_fp16")]; + tensor var_2478 = const()[name = string("op_2478"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_39_cast_fp16 = reshape(shape = var_2478, x = query_39_cast_fp16)[name = string("mh_q_39_cast_fp16")]; + fp16 var_2480_to_fp16 = const()[name = string("op_2480_to_fp16"), val = fp16(0x1p-3)]; + tensor var_2481_cast_fp16 = mul(x = mh_q_39_cast_fp16, y = var_2480_to_fp16)[name = string("op_2481_cast_fp16")]; + tensor var_2482 = const()[name = string("op_2482"), val = tensor([1, 20, 64, -1])]; + tensor var_2483_cast_fp16 = reshape(shape = var_2482, x = key_39_cast_fp16)[name = string("op_2483_cast_fp16")]; + bool mh_w_39_transpose_x_0 = const()[name = string("mh_w_39_transpose_x_0"), val = bool(true)]; + bool mh_w_39_transpose_y_0 = const()[name = string("mh_w_39_transpose_y_0"), val = bool(false)]; + tensor mh_w_39_cast_fp16 = matmul(transpose_x = mh_w_39_transpose_x_0, transpose_y = mh_w_39_transpose_y_0, x = var_2481_cast_fp16, y = var_2483_cast_fp16)[name = string("mh_w_39_cast_fp16")]; + tensor var_2486_cast_fp16 = softmax(axis = var_2424, x = mh_w_39_cast_fp16)[name = string("op_2486_cast_fp16")]; + tensor var_2487 = const()[name = string("op_2487"), val = tensor([1, 20, 64, -1])]; + tensor var_2488_cast_fp16 = reshape(shape = var_2487, x = value_39_cast_fp16)[name = string("op_2488_cast_fp16")]; + bool attn_39_transpose_x_0 = const()[name = string("attn_39_transpose_x_0"), val = bool(false)]; + bool attn_39_transpose_y_0 = const()[name = string("attn_39_transpose_y_0"), val = bool(true)]; + tensor attn_39_cast_fp16 = matmul(transpose_x = attn_39_transpose_x_0, transpose_y = attn_39_transpose_y_0, x = var_2488_cast_fp16, y = var_2486_cast_fp16)[name = string("attn_39_cast_fp16")]; + tensor var_2491 = const()[name = string("op_2491"), val = tensor([1, 1280, 1, -1])]; + tensor input_153_cast_fp16 = reshape(shape = var_2491, x = attn_39_cast_fp16)[name = string("input_153_cast_fp16")]; + string obj_79_pad_type_0 = const()[name = string("obj_79_pad_type_0"), val = string("valid")]; + tensor obj_79_strides_0 = const()[name = string("obj_79_strides_0"), val = tensor([1, 1])]; + tensor obj_79_pad_0 = const()[name = string("obj_79_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_79_dilations_0 = const()[name = string("obj_79_dilations_0"), val = tensor([1, 1])]; + int32 obj_79_groups_0 = const()[name = string("obj_79_groups_0"), val = int32(1)]; + tensor layers_19_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_19_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(772217600)))]; + tensor layers_19_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_19_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(775494464)))]; + tensor obj_79_cast_fp16 = conv(bias = layers_19_self_attn_o_proj_bias_to_fp16, dilations = obj_79_dilations_0, groups = obj_79_groups_0, pad = obj_79_pad_0, pad_type = obj_79_pad_type_0, strides = obj_79_strides_0, weight = layers_19_self_attn_o_proj_weight_to_fp16, x = input_153_cast_fp16)[name = string("obj_79_cast_fp16")]; + tensor inputs_79_cast_fp16 = add(x = inputs_77_cast_fp16, y = obj_79_cast_fp16)[name = string("inputs_79_cast_fp16")]; + tensor out_79_axes_0 = const()[name = string("out_79_axes_0"), val = tensor([1])]; + fp16 var_2509_to_fp16 = const()[name = string("op_2509_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_79_cast_fp16 = layer_norm(axes = out_79_axes_0, epsilon = var_2509_to_fp16, x = inputs_79_cast_fp16)[name = string("out_79_cast_fp16")]; + tensor input_155_gamma_0_to_fp16 = const()[name = string("input_155_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(775497088)))]; + tensor input_155_beta_0_to_fp16 = const()[name = string("input_155_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(775499712)))]; + fp16 input_155_epsilon_0_to_fp16 = const()[name = string("input_155_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor 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 = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_79_cast_fp16)[name = string("input_155_cast_fp16")]; + string input_157_pad_type_0 = const()[name = string("input_157_pad_type_0"), val = string("valid")]; + tensor input_157_strides_0 = const()[name = string("input_157_strides_0"), val = tensor([1, 1])]; + tensor input_157_pad_0 = const()[name = string("input_157_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_157_dilations_0 = const()[name = string("input_157_dilations_0"), val = tensor([1, 1])]; + int32 input_157_groups_0 = const()[name = string("input_157_groups_0"), val = int32(1)]; + tensor layers_19_fc1_weight_to_fp16 = const()[name = string("layers_19_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(775502336)))]; + tensor layers_19_fc1_bias_to_fp16 = const()[name = string("layers_19_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(788609600)))]; + tensor input_157_cast_fp16 = conv(bias = layers_19_fc1_bias_to_fp16, 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_19_fc1_weight_to_fp16, x = input_155_cast_fp16)[name = string("input_157_cast_fp16")]; + string input_159_mode_0 = const()[name = string("input_159_mode_0"), val = string("EXACT")]; + tensor input_159_cast_fp16 = gelu(mode = input_159_mode_0, x = input_157_cast_fp16)[name = string("input_159_cast_fp16")]; + string hidden_states_43_pad_type_0 = const()[name = string("hidden_states_43_pad_type_0"), val = string("valid")]; + tensor hidden_states_43_strides_0 = const()[name = string("hidden_states_43_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_43_pad_0 = const()[name = string("hidden_states_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_43_dilations_0 = const()[name = string("hidden_states_43_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_43_groups_0 = const()[name = string("hidden_states_43_groups_0"), val = int32(1)]; + tensor layers_19_fc2_weight_to_fp16 = const()[name = string("layers_19_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(788619904)))]; + tensor layers_19_fc2_bias_to_fp16 = const()[name = string("layers_19_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(801727168)))]; + tensor hidden_states_43_cast_fp16 = conv(bias = layers_19_fc2_bias_to_fp16, dilations = hidden_states_43_dilations_0, groups = hidden_states_43_groups_0, pad = hidden_states_43_pad_0, pad_type = hidden_states_43_pad_type_0, strides = hidden_states_43_strides_0, weight = layers_19_fc2_weight_to_fp16, x = input_159_cast_fp16)[name = string("hidden_states_43_cast_fp16")]; + tensor inputs_81_cast_fp16 = add(x = inputs_79_cast_fp16, y = hidden_states_43_cast_fp16)[name = string("inputs_81_cast_fp16")]; + int32 var_2542 = const()[name = string("op_2542"), val = int32(3)]; + tensor out_81_axes_0 = const()[name = string("out_81_axes_0"), val = tensor([1])]; + fp16 var_2561_to_fp16 = const()[name = string("op_2561_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_81_cast_fp16 = layer_norm(axes = out_81_axes_0, epsilon = var_2561_to_fp16, x = inputs_81_cast_fp16)[name = string("out_81_cast_fp16")]; + tensor obj_81_gamma_0_to_fp16 = const()[name = string("obj_81_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(801729792)))]; + tensor obj_81_beta_0_to_fp16 = const()[name = string("obj_81_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(801732416)))]; + fp16 obj_81_epsilon_0_to_fp16 = const()[name = string("obj_81_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor obj_81_cast_fp16 = batch_norm(beta = obj_81_beta_0_to_fp16, epsilon = obj_81_epsilon_0_to_fp16, gamma = obj_81_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_81_cast_fp16)[name = string("obj_81_cast_fp16")]; + string query_41_pad_type_0 = const()[name = string("query_41_pad_type_0"), val = string("valid")]; + tensor query_41_strides_0 = const()[name = string("query_41_strides_0"), val = tensor([1, 1])]; + tensor query_41_pad_0 = const()[name = string("query_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_41_dilations_0 = const()[name = string("query_41_dilations_0"), val = tensor([1, 1])]; + int32 query_41_groups_0 = const()[name = string("query_41_groups_0"), val = int32(1)]; + tensor layers_20_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_20_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(801735040)))]; + tensor layers_20_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_20_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(805011904)))]; + tensor query_41_cast_fp16 = conv(bias = layers_20_self_attn_q_proj_bias_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_20_self_attn_q_proj_weight_to_fp16, x = obj_81_cast_fp16)[name = string("query_41_cast_fp16")]; + string key_41_pad_type_0 = const()[name = string("key_41_pad_type_0"), val = string("valid")]; + tensor key_41_strides_0 = const()[name = string("key_41_strides_0"), val = tensor([1, 1])]; + tensor key_41_pad_0 = const()[name = string("key_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_41_dilations_0 = const()[name = string("key_41_dilations_0"), val = tensor([1, 1])]; + int32 key_41_groups_0 = const()[name = string("key_41_groups_0"), val = int32(1)]; + tensor layers_20_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_20_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(805014528)))]; + tensor key_41_cast_fp16 = conv(dilations = key_41_dilations_0, groups = key_41_groups_0, pad = key_41_pad_0, pad_type = key_41_pad_type_0, strides = key_41_strides_0, weight = layers_20_self_attn_k_proj_weight_to_fp16, x = obj_81_cast_fp16)[name = string("key_41_cast_fp16")]; + string value_41_pad_type_0 = const()[name = string("value_41_pad_type_0"), val = string("valid")]; + tensor value_41_strides_0 = const()[name = string("value_41_strides_0"), val = tensor([1, 1])]; + tensor value_41_pad_0 = const()[name = string("value_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_41_dilations_0 = const()[name = string("value_41_dilations_0"), val = tensor([1, 1])]; + int32 value_41_groups_0 = const()[name = string("value_41_groups_0"), val = int32(1)]; + tensor layers_20_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_20_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(808291392)))]; + tensor layers_20_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_20_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(811568256)))]; + tensor value_41_cast_fp16 = conv(bias = layers_20_self_attn_v_proj_bias_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_81_cast_fp16)[name = string("value_41_cast_fp16")]; + tensor var_2596 = const()[name = string("op_2596"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_41_cast_fp16 = reshape(shape = var_2596, x = query_41_cast_fp16)[name = string("mh_q_41_cast_fp16")]; + fp16 var_2598_to_fp16 = const()[name = string("op_2598_to_fp16"), val = fp16(0x1p-3)]; + tensor var_2599_cast_fp16 = mul(x = mh_q_41_cast_fp16, y = var_2598_to_fp16)[name = string("op_2599_cast_fp16")]; + tensor var_2600 = const()[name = string("op_2600"), val = tensor([1, 20, 64, -1])]; + tensor var_2601_cast_fp16 = reshape(shape = var_2600, x = key_41_cast_fp16)[name = string("op_2601_cast_fp16")]; + bool mh_w_41_transpose_x_0 = const()[name = string("mh_w_41_transpose_x_0"), val = bool(true)]; + bool mh_w_41_transpose_y_0 = const()[name = string("mh_w_41_transpose_y_0"), val = bool(false)]; + tensor mh_w_41_cast_fp16 = matmul(transpose_x = mh_w_41_transpose_x_0, transpose_y = mh_w_41_transpose_y_0, x = var_2599_cast_fp16, y = var_2601_cast_fp16)[name = string("mh_w_41_cast_fp16")]; + tensor var_2604_cast_fp16 = softmax(axis = var_2542, x = mh_w_41_cast_fp16)[name = string("op_2604_cast_fp16")]; + tensor var_2605 = const()[name = string("op_2605"), val = tensor([1, 20, 64, -1])]; + tensor var_2606_cast_fp16 = reshape(shape = var_2605, x = value_41_cast_fp16)[name = string("op_2606_cast_fp16")]; + bool attn_41_transpose_x_0 = const()[name = string("attn_41_transpose_x_0"), val = bool(false)]; + bool attn_41_transpose_y_0 = const()[name = string("attn_41_transpose_y_0"), val = bool(true)]; + tensor attn_41_cast_fp16 = matmul(transpose_x = attn_41_transpose_x_0, transpose_y = attn_41_transpose_y_0, x = var_2606_cast_fp16, y = var_2604_cast_fp16)[name = string("attn_41_cast_fp16")]; + tensor var_2609 = const()[name = string("op_2609"), val = tensor([1, 1280, 1, -1])]; + tensor input_161_cast_fp16 = reshape(shape = var_2609, x = attn_41_cast_fp16)[name = string("input_161_cast_fp16")]; + string obj_83_pad_type_0 = const()[name = string("obj_83_pad_type_0"), val = string("valid")]; + tensor obj_83_strides_0 = const()[name = string("obj_83_strides_0"), val = tensor([1, 1])]; + tensor obj_83_pad_0 = const()[name = string("obj_83_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_83_dilations_0 = const()[name = string("obj_83_dilations_0"), val = tensor([1, 1])]; + int32 obj_83_groups_0 = const()[name = string("obj_83_groups_0"), val = int32(1)]; + tensor layers_20_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_20_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(811570880)))]; + tensor layers_20_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_20_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(814847744)))]; + tensor obj_83_cast_fp16 = conv(bias = layers_20_self_attn_o_proj_bias_to_fp16, dilations = obj_83_dilations_0, groups = obj_83_groups_0, pad = obj_83_pad_0, pad_type = obj_83_pad_type_0, strides = obj_83_strides_0, weight = layers_20_self_attn_o_proj_weight_to_fp16, x = input_161_cast_fp16)[name = string("obj_83_cast_fp16")]; + tensor inputs_83_cast_fp16 = add(x = inputs_81_cast_fp16, y = obj_83_cast_fp16)[name = string("inputs_83_cast_fp16")]; + tensor out_83_axes_0 = const()[name = string("out_83_axes_0"), val = tensor([1])]; + fp16 var_2627_to_fp16 = const()[name = string("op_2627_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_83_cast_fp16 = layer_norm(axes = out_83_axes_0, epsilon = var_2627_to_fp16, x = inputs_83_cast_fp16)[name = string("out_83_cast_fp16")]; + tensor input_163_gamma_0_to_fp16 = const()[name = string("input_163_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(814850368)))]; + tensor input_163_beta_0_to_fp16 = const()[name = string("input_163_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(814852992)))]; + fp16 input_163_epsilon_0_to_fp16 = const()[name = string("input_163_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor input_163_cast_fp16 = batch_norm(beta = input_163_beta_0_to_fp16, epsilon = input_163_epsilon_0_to_fp16, gamma = input_163_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_83_cast_fp16)[name = string("input_163_cast_fp16")]; + string input_165_pad_type_0 = const()[name = string("input_165_pad_type_0"), val = string("valid")]; + tensor input_165_strides_0 = const()[name = string("input_165_strides_0"), val = tensor([1, 1])]; + tensor input_165_pad_0 = const()[name = string("input_165_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_165_dilations_0 = const()[name = string("input_165_dilations_0"), val = tensor([1, 1])]; + int32 input_165_groups_0 = const()[name = string("input_165_groups_0"), val = int32(1)]; + tensor layers_20_fc1_weight_to_fp16 = const()[name = string("layers_20_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(814855616)))]; + tensor layers_20_fc1_bias_to_fp16 = const()[name = string("layers_20_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(827962880)))]; + tensor input_165_cast_fp16 = conv(bias = layers_20_fc1_bias_to_fp16, dilations = input_165_dilations_0, groups = input_165_groups_0, pad = input_165_pad_0, pad_type = input_165_pad_type_0, strides = input_165_strides_0, weight = layers_20_fc1_weight_to_fp16, x = input_163_cast_fp16)[name = string("input_165_cast_fp16")]; + string input_167_mode_0 = const()[name = string("input_167_mode_0"), val = string("EXACT")]; + tensor input_167_cast_fp16 = gelu(mode = input_167_mode_0, x = input_165_cast_fp16)[name = string("input_167_cast_fp16")]; + string hidden_states_45_pad_type_0 = const()[name = string("hidden_states_45_pad_type_0"), val = string("valid")]; + tensor hidden_states_45_strides_0 = const()[name = string("hidden_states_45_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_45_pad_0 = const()[name = string("hidden_states_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_45_dilations_0 = const()[name = string("hidden_states_45_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_45_groups_0 = const()[name = string("hidden_states_45_groups_0"), val = int32(1)]; + tensor layers_20_fc2_weight_to_fp16 = const()[name = string("layers_20_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(827973184)))]; + tensor layers_20_fc2_bias_to_fp16 = const()[name = string("layers_20_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(841080448)))]; + tensor hidden_states_45_cast_fp16 = conv(bias = layers_20_fc2_bias_to_fp16, dilations = hidden_states_45_dilations_0, groups = hidden_states_45_groups_0, pad = hidden_states_45_pad_0, pad_type = hidden_states_45_pad_type_0, strides = hidden_states_45_strides_0, weight = layers_20_fc2_weight_to_fp16, x = input_167_cast_fp16)[name = string("hidden_states_45_cast_fp16")]; + tensor inputs_85_cast_fp16 = add(x = inputs_83_cast_fp16, y = hidden_states_45_cast_fp16)[name = string("inputs_85_cast_fp16")]; + int32 var_2660 = const()[name = string("op_2660"), val = int32(3)]; + tensor out_85_axes_0 = const()[name = string("out_85_axes_0"), val = tensor([1])]; + fp16 var_2679_to_fp16 = const()[name = string("op_2679_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_85_cast_fp16 = layer_norm(axes = out_85_axes_0, epsilon = var_2679_to_fp16, x = inputs_85_cast_fp16)[name = string("out_85_cast_fp16")]; + tensor obj_85_gamma_0_to_fp16 = const()[name = string("obj_85_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(841083072)))]; + tensor obj_85_beta_0_to_fp16 = const()[name = string("obj_85_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(841085696)))]; + fp16 obj_85_epsilon_0_to_fp16 = const()[name = string("obj_85_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor obj_85_cast_fp16 = batch_norm(beta = obj_85_beta_0_to_fp16, epsilon = obj_85_epsilon_0_to_fp16, gamma = obj_85_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_85_cast_fp16)[name = string("obj_85_cast_fp16")]; + string query_43_pad_type_0 = const()[name = string("query_43_pad_type_0"), val = string("valid")]; + tensor query_43_strides_0 = const()[name = string("query_43_strides_0"), val = tensor([1, 1])]; + tensor query_43_pad_0 = const()[name = string("query_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_43_dilations_0 = const()[name = string("query_43_dilations_0"), val = tensor([1, 1])]; + int32 query_43_groups_0 = const()[name = string("query_43_groups_0"), val = int32(1)]; + tensor layers_21_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_21_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(841088320)))]; + tensor layers_21_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_21_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(844365184)))]; + tensor query_43_cast_fp16 = conv(bias = layers_21_self_attn_q_proj_bias_to_fp16, dilations = query_43_dilations_0, groups = query_43_groups_0, pad = query_43_pad_0, pad_type = query_43_pad_type_0, strides = query_43_strides_0, weight = layers_21_self_attn_q_proj_weight_to_fp16, x = obj_85_cast_fp16)[name = string("query_43_cast_fp16")]; + string key_43_pad_type_0 = const()[name = string("key_43_pad_type_0"), val = string("valid")]; + tensor key_43_strides_0 = const()[name = string("key_43_strides_0"), val = tensor([1, 1])]; + tensor key_43_pad_0 = const()[name = string("key_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_43_dilations_0 = const()[name = string("key_43_dilations_0"), val = tensor([1, 1])]; + int32 key_43_groups_0 = const()[name = string("key_43_groups_0"), val = int32(1)]; + tensor layers_21_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_21_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(844367808)))]; + tensor key_43_cast_fp16 = conv(dilations = key_43_dilations_0, groups = key_43_groups_0, pad = key_43_pad_0, pad_type = key_43_pad_type_0, strides = key_43_strides_0, weight = layers_21_self_attn_k_proj_weight_to_fp16, x = obj_85_cast_fp16)[name = string("key_43_cast_fp16")]; + string value_43_pad_type_0 = const()[name = string("value_43_pad_type_0"), val = string("valid")]; + tensor value_43_strides_0 = const()[name = string("value_43_strides_0"), val = tensor([1, 1])]; + tensor value_43_pad_0 = const()[name = string("value_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_43_dilations_0 = const()[name = string("value_43_dilations_0"), val = tensor([1, 1])]; + int32 value_43_groups_0 = const()[name = string("value_43_groups_0"), val = int32(1)]; + tensor layers_21_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_21_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(847644672)))]; + tensor layers_21_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_21_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(850921536)))]; + tensor value_43_cast_fp16 = conv(bias = layers_21_self_attn_v_proj_bias_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_85_cast_fp16)[name = string("value_43_cast_fp16")]; + tensor var_2714 = const()[name = string("op_2714"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_43_cast_fp16 = reshape(shape = var_2714, x = query_43_cast_fp16)[name = string("mh_q_43_cast_fp16")]; + fp16 var_2716_to_fp16 = const()[name = string("op_2716_to_fp16"), val = fp16(0x1p-3)]; + tensor var_2717_cast_fp16 = mul(x = mh_q_43_cast_fp16, y = var_2716_to_fp16)[name = string("op_2717_cast_fp16")]; + tensor var_2718 = const()[name = string("op_2718"), val = tensor([1, 20, 64, -1])]; + tensor var_2719_cast_fp16 = reshape(shape = var_2718, x = key_43_cast_fp16)[name = string("op_2719_cast_fp16")]; + bool mh_w_43_transpose_x_0 = const()[name = string("mh_w_43_transpose_x_0"), val = bool(true)]; + bool mh_w_43_transpose_y_0 = const()[name = string("mh_w_43_transpose_y_0"), val = bool(false)]; + tensor mh_w_43_cast_fp16 = matmul(transpose_x = mh_w_43_transpose_x_0, transpose_y = mh_w_43_transpose_y_0, x = var_2717_cast_fp16, y = var_2719_cast_fp16)[name = string("mh_w_43_cast_fp16")]; + tensor var_2722_cast_fp16 = softmax(axis = var_2660, x = mh_w_43_cast_fp16)[name = string("op_2722_cast_fp16")]; + tensor var_2723 = const()[name = string("op_2723"), val = tensor([1, 20, 64, -1])]; + tensor var_2724_cast_fp16 = reshape(shape = var_2723, x = value_43_cast_fp16)[name = string("op_2724_cast_fp16")]; + bool attn_43_transpose_x_0 = const()[name = string("attn_43_transpose_x_0"), val = bool(false)]; + bool attn_43_transpose_y_0 = const()[name = string("attn_43_transpose_y_0"), val = bool(true)]; + tensor attn_43_cast_fp16 = matmul(transpose_x = attn_43_transpose_x_0, transpose_y = attn_43_transpose_y_0, x = var_2724_cast_fp16, y = var_2722_cast_fp16)[name = string("attn_43_cast_fp16")]; + tensor var_2727 = const()[name = string("op_2727"), val = tensor([1, 1280, 1, -1])]; + tensor input_169_cast_fp16 = reshape(shape = var_2727, x = attn_43_cast_fp16)[name = string("input_169_cast_fp16")]; + string obj_87_pad_type_0 = const()[name = string("obj_87_pad_type_0"), val = string("valid")]; + tensor obj_87_strides_0 = const()[name = string("obj_87_strides_0"), val = tensor([1, 1])]; + tensor obj_87_pad_0 = const()[name = string("obj_87_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_87_dilations_0 = const()[name = string("obj_87_dilations_0"), val = tensor([1, 1])]; + int32 obj_87_groups_0 = const()[name = string("obj_87_groups_0"), val = int32(1)]; + tensor layers_21_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_21_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(850924160)))]; + tensor layers_21_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_21_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(854201024)))]; + tensor obj_87_cast_fp16 = conv(bias = layers_21_self_attn_o_proj_bias_to_fp16, dilations = obj_87_dilations_0, groups = obj_87_groups_0, pad = obj_87_pad_0, pad_type = obj_87_pad_type_0, strides = obj_87_strides_0, weight = layers_21_self_attn_o_proj_weight_to_fp16, x = input_169_cast_fp16)[name = string("obj_87_cast_fp16")]; + tensor inputs_87_cast_fp16 = add(x = inputs_85_cast_fp16, y = obj_87_cast_fp16)[name = string("inputs_87_cast_fp16")]; + tensor out_87_axes_0 = const()[name = string("out_87_axes_0"), val = tensor([1])]; + fp16 var_2745_to_fp16 = const()[name = string("op_2745_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_87_cast_fp16 = layer_norm(axes = out_87_axes_0, epsilon = var_2745_to_fp16, x = inputs_87_cast_fp16)[name = string("out_87_cast_fp16")]; + tensor input_171_gamma_0_to_fp16 = const()[name = string("input_171_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(854203648)))]; + tensor input_171_beta_0_to_fp16 = const()[name = string("input_171_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(854206272)))]; + fp16 input_171_epsilon_0_to_fp16 = const()[name = string("input_171_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor 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 = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_87_cast_fp16)[name = string("input_171_cast_fp16")]; + string input_173_pad_type_0 = const()[name = string("input_173_pad_type_0"), val = string("valid")]; + tensor input_173_strides_0 = const()[name = string("input_173_strides_0"), val = tensor([1, 1])]; + tensor input_173_pad_0 = const()[name = string("input_173_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_173_dilations_0 = const()[name = string("input_173_dilations_0"), val = tensor([1, 1])]; + int32 input_173_groups_0 = const()[name = string("input_173_groups_0"), val = int32(1)]; + tensor layers_21_fc1_weight_to_fp16 = const()[name = string("layers_21_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(854208896)))]; + tensor layers_21_fc1_bias_to_fp16 = const()[name = string("layers_21_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(867316160)))]; + tensor input_173_cast_fp16 = conv(bias = layers_21_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_21_fc1_weight_to_fp16, x = input_171_cast_fp16)[name = string("input_173_cast_fp16")]; + string input_175_mode_0 = const()[name = string("input_175_mode_0"), val = string("EXACT")]; + tensor input_175_cast_fp16 = gelu(mode = input_175_mode_0, x = input_173_cast_fp16)[name = string("input_175_cast_fp16")]; + string hidden_states_47_pad_type_0 = const()[name = string("hidden_states_47_pad_type_0"), val = string("valid")]; + tensor hidden_states_47_strides_0 = const()[name = string("hidden_states_47_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_47_pad_0 = const()[name = string("hidden_states_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_47_dilations_0 = const()[name = string("hidden_states_47_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_47_groups_0 = const()[name = string("hidden_states_47_groups_0"), val = int32(1)]; + tensor layers_21_fc2_weight_to_fp16 = const()[name = string("layers_21_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(867326464)))]; + tensor layers_21_fc2_bias_to_fp16 = const()[name = string("layers_21_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(880433728)))]; + tensor hidden_states_47_cast_fp16 = conv(bias = layers_21_fc2_bias_to_fp16, dilations = hidden_states_47_dilations_0, groups = hidden_states_47_groups_0, pad = hidden_states_47_pad_0, pad_type = hidden_states_47_pad_type_0, strides = hidden_states_47_strides_0, weight = layers_21_fc2_weight_to_fp16, x = input_175_cast_fp16)[name = string("hidden_states_47_cast_fp16")]; + tensor inputs_89_cast_fp16 = add(x = inputs_87_cast_fp16, y = hidden_states_47_cast_fp16)[name = string("inputs_89_cast_fp16")]; + int32 var_2778 = const()[name = string("op_2778"), val = int32(3)]; + tensor out_89_axes_0 = const()[name = string("out_89_axes_0"), val = tensor([1])]; + fp16 var_2797_to_fp16 = const()[name = string("op_2797_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_89_cast_fp16 = layer_norm(axes = out_89_axes_0, epsilon = var_2797_to_fp16, x = inputs_89_cast_fp16)[name = string("out_89_cast_fp16")]; + tensor obj_89_gamma_0_to_fp16 = const()[name = string("obj_89_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(880436352)))]; + tensor obj_89_beta_0_to_fp16 = const()[name = string("obj_89_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(880438976)))]; + fp16 obj_89_epsilon_0_to_fp16 = const()[name = string("obj_89_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor obj_89_cast_fp16 = batch_norm(beta = obj_89_beta_0_to_fp16, epsilon = obj_89_epsilon_0_to_fp16, gamma = obj_89_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_89_cast_fp16)[name = string("obj_89_cast_fp16")]; + string query_45_pad_type_0 = const()[name = string("query_45_pad_type_0"), val = string("valid")]; + tensor query_45_strides_0 = const()[name = string("query_45_strides_0"), val = tensor([1, 1])]; + tensor query_45_pad_0 = const()[name = string("query_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_45_dilations_0 = const()[name = string("query_45_dilations_0"), val = tensor([1, 1])]; + int32 query_45_groups_0 = const()[name = string("query_45_groups_0"), val = int32(1)]; + tensor layers_22_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_22_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(880441600)))]; + tensor layers_22_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_22_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(883718464)))]; + tensor query_45_cast_fp16 = conv(bias = layers_22_self_attn_q_proj_bias_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_22_self_attn_q_proj_weight_to_fp16, x = obj_89_cast_fp16)[name = string("query_45_cast_fp16")]; + string key_45_pad_type_0 = const()[name = string("key_45_pad_type_0"), val = string("valid")]; + tensor key_45_strides_0 = const()[name = string("key_45_strides_0"), val = tensor([1, 1])]; + tensor key_45_pad_0 = const()[name = string("key_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_45_dilations_0 = const()[name = string("key_45_dilations_0"), val = tensor([1, 1])]; + int32 key_45_groups_0 = const()[name = string("key_45_groups_0"), val = int32(1)]; + tensor layers_22_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_22_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(883721088)))]; + tensor key_45_cast_fp16 = conv(dilations = key_45_dilations_0, groups = key_45_groups_0, pad = key_45_pad_0, pad_type = key_45_pad_type_0, strides = key_45_strides_0, weight = layers_22_self_attn_k_proj_weight_to_fp16, x = obj_89_cast_fp16)[name = string("key_45_cast_fp16")]; + string value_45_pad_type_0 = const()[name = string("value_45_pad_type_0"), val = string("valid")]; + tensor value_45_strides_0 = const()[name = string("value_45_strides_0"), val = tensor([1, 1])]; + tensor value_45_pad_0 = const()[name = string("value_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_45_dilations_0 = const()[name = string("value_45_dilations_0"), val = tensor([1, 1])]; + int32 value_45_groups_0 = const()[name = string("value_45_groups_0"), val = int32(1)]; + tensor layers_22_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_22_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(886997952)))]; + tensor layers_22_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_22_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(890274816)))]; + tensor value_45_cast_fp16 = conv(bias = layers_22_self_attn_v_proj_bias_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_89_cast_fp16)[name = string("value_45_cast_fp16")]; + tensor var_2832 = const()[name = string("op_2832"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_45_cast_fp16 = reshape(shape = var_2832, x = query_45_cast_fp16)[name = string("mh_q_45_cast_fp16")]; + fp16 var_2834_to_fp16 = const()[name = string("op_2834_to_fp16"), val = fp16(0x1p-3)]; + tensor var_2835_cast_fp16 = mul(x = mh_q_45_cast_fp16, y = var_2834_to_fp16)[name = string("op_2835_cast_fp16")]; + tensor var_2836 = const()[name = string("op_2836"), val = tensor([1, 20, 64, -1])]; + tensor var_2837_cast_fp16 = reshape(shape = var_2836, x = key_45_cast_fp16)[name = string("op_2837_cast_fp16")]; + bool mh_w_45_transpose_x_0 = const()[name = string("mh_w_45_transpose_x_0"), val = bool(true)]; + bool mh_w_45_transpose_y_0 = const()[name = string("mh_w_45_transpose_y_0"), val = bool(false)]; + tensor mh_w_45_cast_fp16 = matmul(transpose_x = mh_w_45_transpose_x_0, transpose_y = mh_w_45_transpose_y_0, x = var_2835_cast_fp16, y = var_2837_cast_fp16)[name = string("mh_w_45_cast_fp16")]; + tensor var_2840_cast_fp16 = softmax(axis = var_2778, x = mh_w_45_cast_fp16)[name = string("op_2840_cast_fp16")]; + tensor var_2841 = const()[name = string("op_2841"), val = tensor([1, 20, 64, -1])]; + tensor var_2842_cast_fp16 = reshape(shape = var_2841, x = value_45_cast_fp16)[name = string("op_2842_cast_fp16")]; + bool attn_45_transpose_x_0 = const()[name = string("attn_45_transpose_x_0"), val = bool(false)]; + bool attn_45_transpose_y_0 = const()[name = string("attn_45_transpose_y_0"), val = bool(true)]; + tensor attn_45_cast_fp16 = matmul(transpose_x = attn_45_transpose_x_0, transpose_y = attn_45_transpose_y_0, x = var_2842_cast_fp16, y = var_2840_cast_fp16)[name = string("attn_45_cast_fp16")]; + tensor var_2845 = const()[name = string("op_2845"), val = tensor([1, 1280, 1, -1])]; + tensor input_177_cast_fp16 = reshape(shape = var_2845, x = attn_45_cast_fp16)[name = string("input_177_cast_fp16")]; + string obj_91_pad_type_0 = const()[name = string("obj_91_pad_type_0"), val = string("valid")]; + tensor obj_91_strides_0 = const()[name = string("obj_91_strides_0"), val = tensor([1, 1])]; + tensor obj_91_pad_0 = const()[name = string("obj_91_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_91_dilations_0 = const()[name = string("obj_91_dilations_0"), val = tensor([1, 1])]; + int32 obj_91_groups_0 = const()[name = string("obj_91_groups_0"), val = int32(1)]; + tensor layers_22_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_22_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(890277440)))]; + tensor layers_22_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_22_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(893554304)))]; + tensor obj_91_cast_fp16 = conv(bias = layers_22_self_attn_o_proj_bias_to_fp16, dilations = obj_91_dilations_0, groups = obj_91_groups_0, pad = obj_91_pad_0, pad_type = obj_91_pad_type_0, strides = obj_91_strides_0, weight = layers_22_self_attn_o_proj_weight_to_fp16, x = input_177_cast_fp16)[name = string("obj_91_cast_fp16")]; + tensor inputs_91_cast_fp16 = add(x = inputs_89_cast_fp16, y = obj_91_cast_fp16)[name = string("inputs_91_cast_fp16")]; + tensor out_91_axes_0 = const()[name = string("out_91_axes_0"), val = tensor([1])]; + fp16 var_2863_to_fp16 = const()[name = string("op_2863_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_91_cast_fp16 = layer_norm(axes = out_91_axes_0, epsilon = var_2863_to_fp16, x = inputs_91_cast_fp16)[name = string("out_91_cast_fp16")]; + tensor input_179_gamma_0_to_fp16 = const()[name = string("input_179_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(893556928)))]; + tensor input_179_beta_0_to_fp16 = const()[name = string("input_179_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(893559552)))]; + fp16 input_179_epsilon_0_to_fp16 = const()[name = string("input_179_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor 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 = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_91_cast_fp16)[name = string("input_179_cast_fp16")]; + string input_181_pad_type_0 = const()[name = string("input_181_pad_type_0"), val = string("valid")]; + tensor input_181_strides_0 = const()[name = string("input_181_strides_0"), val = tensor([1, 1])]; + tensor input_181_pad_0 = const()[name = string("input_181_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_181_dilations_0 = const()[name = string("input_181_dilations_0"), val = tensor([1, 1])]; + int32 input_181_groups_0 = const()[name = string("input_181_groups_0"), val = int32(1)]; + tensor layers_22_fc1_weight_to_fp16 = const()[name = string("layers_22_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(893562176)))]; + tensor layers_22_fc1_bias_to_fp16 = const()[name = string("layers_22_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(906669440)))]; + tensor input_181_cast_fp16 = conv(bias = layers_22_fc1_bias_to_fp16, 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_22_fc1_weight_to_fp16, x = input_179_cast_fp16)[name = string("input_181_cast_fp16")]; + string input_183_mode_0 = const()[name = string("input_183_mode_0"), val = string("EXACT")]; + tensor input_183_cast_fp16 = gelu(mode = input_183_mode_0, x = input_181_cast_fp16)[name = string("input_183_cast_fp16")]; + string hidden_states_49_pad_type_0 = const()[name = string("hidden_states_49_pad_type_0"), val = string("valid")]; + tensor hidden_states_49_strides_0 = const()[name = string("hidden_states_49_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_49_pad_0 = const()[name = string("hidden_states_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_49_dilations_0 = const()[name = string("hidden_states_49_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_49_groups_0 = const()[name = string("hidden_states_49_groups_0"), val = int32(1)]; + tensor layers_22_fc2_weight_to_fp16 = const()[name = string("layers_22_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(906679744)))]; + tensor layers_22_fc2_bias_to_fp16 = const()[name = string("layers_22_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(919787008)))]; + tensor hidden_states_49_cast_fp16 = conv(bias = layers_22_fc2_bias_to_fp16, dilations = hidden_states_49_dilations_0, groups = hidden_states_49_groups_0, pad = hidden_states_49_pad_0, pad_type = hidden_states_49_pad_type_0, strides = hidden_states_49_strides_0, weight = layers_22_fc2_weight_to_fp16, x = input_183_cast_fp16)[name = string("hidden_states_49_cast_fp16")]; + tensor inputs_93_cast_fp16 = add(x = inputs_91_cast_fp16, y = hidden_states_49_cast_fp16)[name = string("inputs_93_cast_fp16")]; + int32 var_2896 = const()[name = string("op_2896"), val = int32(3)]; + tensor out_93_axes_0 = const()[name = string("out_93_axes_0"), val = tensor([1])]; + fp16 var_2915_to_fp16 = const()[name = string("op_2915_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_93_cast_fp16 = layer_norm(axes = out_93_axes_0, epsilon = var_2915_to_fp16, x = inputs_93_cast_fp16)[name = string("out_93_cast_fp16")]; + tensor obj_93_gamma_0_to_fp16 = const()[name = string("obj_93_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(919789632)))]; + tensor obj_93_beta_0_to_fp16 = const()[name = string("obj_93_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(919792256)))]; + fp16 obj_93_epsilon_0_to_fp16 = const()[name = string("obj_93_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor obj_93_cast_fp16 = batch_norm(beta = obj_93_beta_0_to_fp16, epsilon = obj_93_epsilon_0_to_fp16, gamma = obj_93_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_93_cast_fp16)[name = string("obj_93_cast_fp16")]; + string query_47_pad_type_0 = const()[name = string("query_47_pad_type_0"), val = string("valid")]; + tensor query_47_strides_0 = const()[name = string("query_47_strides_0"), val = tensor([1, 1])]; + tensor query_47_pad_0 = const()[name = string("query_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_47_dilations_0 = const()[name = string("query_47_dilations_0"), val = tensor([1, 1])]; + int32 query_47_groups_0 = const()[name = string("query_47_groups_0"), val = int32(1)]; + tensor layers_23_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_23_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(919794880)))]; + tensor layers_23_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_23_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(923071744)))]; + tensor query_47_cast_fp16 = conv(bias = layers_23_self_attn_q_proj_bias_to_fp16, dilations = query_47_dilations_0, groups = query_47_groups_0, pad = query_47_pad_0, pad_type = query_47_pad_type_0, strides = query_47_strides_0, weight = layers_23_self_attn_q_proj_weight_to_fp16, x = obj_93_cast_fp16)[name = string("query_47_cast_fp16")]; + string key_47_pad_type_0 = const()[name = string("key_47_pad_type_0"), val = string("valid")]; + tensor key_47_strides_0 = const()[name = string("key_47_strides_0"), val = tensor([1, 1])]; + tensor key_47_pad_0 = const()[name = string("key_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_47_dilations_0 = const()[name = string("key_47_dilations_0"), val = tensor([1, 1])]; + int32 key_47_groups_0 = const()[name = string("key_47_groups_0"), val = int32(1)]; + tensor layers_23_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_23_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(923074368)))]; + tensor key_47_cast_fp16 = conv(dilations = key_47_dilations_0, groups = key_47_groups_0, pad = key_47_pad_0, pad_type = key_47_pad_type_0, strides = key_47_strides_0, weight = layers_23_self_attn_k_proj_weight_to_fp16, x = obj_93_cast_fp16)[name = string("key_47_cast_fp16")]; + string value_47_pad_type_0 = const()[name = string("value_47_pad_type_0"), val = string("valid")]; + tensor value_47_strides_0 = const()[name = string("value_47_strides_0"), val = tensor([1, 1])]; + tensor value_47_pad_0 = const()[name = string("value_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_47_dilations_0 = const()[name = string("value_47_dilations_0"), val = tensor([1, 1])]; + int32 value_47_groups_0 = const()[name = string("value_47_groups_0"), val = int32(1)]; + tensor layers_23_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_23_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(926351232)))]; + tensor layers_23_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_23_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(929628096)))]; + tensor value_47_cast_fp16 = conv(bias = layers_23_self_attn_v_proj_bias_to_fp16, dilations = value_47_dilations_0, groups = value_47_groups_0, pad = value_47_pad_0, pad_type = value_47_pad_type_0, strides = value_47_strides_0, weight = layers_23_self_attn_v_proj_weight_to_fp16, x = obj_93_cast_fp16)[name = string("value_47_cast_fp16")]; + tensor var_2950 = const()[name = string("op_2950"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_47_cast_fp16 = reshape(shape = var_2950, x = query_47_cast_fp16)[name = string("mh_q_47_cast_fp16")]; + fp16 var_2952_to_fp16 = const()[name = string("op_2952_to_fp16"), val = fp16(0x1p-3)]; + tensor var_2953_cast_fp16 = mul(x = mh_q_47_cast_fp16, y = var_2952_to_fp16)[name = string("op_2953_cast_fp16")]; + tensor var_2954 = const()[name = string("op_2954"), val = tensor([1, 20, 64, -1])]; + tensor var_2955_cast_fp16 = reshape(shape = var_2954, x = key_47_cast_fp16)[name = string("op_2955_cast_fp16")]; + bool mh_w_47_transpose_x_0 = const()[name = string("mh_w_47_transpose_x_0"), val = bool(true)]; + bool mh_w_47_transpose_y_0 = const()[name = string("mh_w_47_transpose_y_0"), val = bool(false)]; + tensor mh_w_47_cast_fp16 = matmul(transpose_x = mh_w_47_transpose_x_0, transpose_y = mh_w_47_transpose_y_0, x = var_2953_cast_fp16, y = var_2955_cast_fp16)[name = string("mh_w_47_cast_fp16")]; + tensor var_2958_cast_fp16 = softmax(axis = var_2896, x = mh_w_47_cast_fp16)[name = string("op_2958_cast_fp16")]; + tensor var_2959 = const()[name = string("op_2959"), val = tensor([1, 20, 64, -1])]; + tensor var_2960_cast_fp16 = reshape(shape = var_2959, x = value_47_cast_fp16)[name = string("op_2960_cast_fp16")]; + bool attn_47_transpose_x_0 = const()[name = string("attn_47_transpose_x_0"), val = bool(false)]; + bool attn_47_transpose_y_0 = const()[name = string("attn_47_transpose_y_0"), val = bool(true)]; + tensor attn_47_cast_fp16 = matmul(transpose_x = attn_47_transpose_x_0, transpose_y = attn_47_transpose_y_0, x = var_2960_cast_fp16, y = var_2958_cast_fp16)[name = string("attn_47_cast_fp16")]; + tensor var_2963 = const()[name = string("op_2963"), val = tensor([1, 1280, 1, -1])]; + tensor input_185_cast_fp16 = reshape(shape = var_2963, x = attn_47_cast_fp16)[name = string("input_185_cast_fp16")]; + string obj_95_pad_type_0 = const()[name = string("obj_95_pad_type_0"), val = string("valid")]; + tensor obj_95_strides_0 = const()[name = string("obj_95_strides_0"), val = tensor([1, 1])]; + tensor obj_95_pad_0 = const()[name = string("obj_95_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_95_dilations_0 = const()[name = string("obj_95_dilations_0"), val = tensor([1, 1])]; + int32 obj_95_groups_0 = const()[name = string("obj_95_groups_0"), val = int32(1)]; + tensor layers_23_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_23_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(929630720)))]; + tensor layers_23_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_23_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(932907584)))]; + tensor obj_95_cast_fp16 = conv(bias = layers_23_self_attn_o_proj_bias_to_fp16, dilations = obj_95_dilations_0, groups = obj_95_groups_0, pad = obj_95_pad_0, pad_type = obj_95_pad_type_0, strides = obj_95_strides_0, weight = layers_23_self_attn_o_proj_weight_to_fp16, x = input_185_cast_fp16)[name = string("obj_95_cast_fp16")]; + tensor inputs_95_cast_fp16 = add(x = inputs_93_cast_fp16, y = obj_95_cast_fp16)[name = string("inputs_95_cast_fp16")]; + tensor out_95_axes_0 = const()[name = string("out_95_axes_0"), val = tensor([1])]; + fp16 var_2981_to_fp16 = const()[name = string("op_2981_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_95_cast_fp16 = layer_norm(axes = out_95_axes_0, epsilon = var_2981_to_fp16, x = inputs_95_cast_fp16)[name = string("out_95_cast_fp16")]; + tensor input_187_gamma_0_to_fp16 = const()[name = string("input_187_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(932910208)))]; + tensor input_187_beta_0_to_fp16 = const()[name = string("input_187_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(932912832)))]; + fp16 input_187_epsilon_0_to_fp16 = const()[name = string("input_187_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor input_187_cast_fp16 = batch_norm(beta = input_187_beta_0_to_fp16, epsilon = input_187_epsilon_0_to_fp16, gamma = input_187_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_95_cast_fp16)[name = string("input_187_cast_fp16")]; + string input_189_pad_type_0 = const()[name = string("input_189_pad_type_0"), val = string("valid")]; + tensor input_189_strides_0 = const()[name = string("input_189_strides_0"), val = tensor([1, 1])]; + tensor input_189_pad_0 = const()[name = string("input_189_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_189_dilations_0 = const()[name = string("input_189_dilations_0"), val = tensor([1, 1])]; + int32 input_189_groups_0 = const()[name = string("input_189_groups_0"), val = int32(1)]; + tensor layers_23_fc1_weight_to_fp16 = const()[name = string("layers_23_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(932915456)))]; + tensor layers_23_fc1_bias_to_fp16 = const()[name = string("layers_23_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(946022720)))]; + tensor input_189_cast_fp16 = conv(bias = layers_23_fc1_bias_to_fp16, dilations = input_189_dilations_0, groups = input_189_groups_0, pad = input_189_pad_0, pad_type = input_189_pad_type_0, strides = input_189_strides_0, weight = layers_23_fc1_weight_to_fp16, x = input_187_cast_fp16)[name = string("input_189_cast_fp16")]; + string input_191_mode_0 = const()[name = string("input_191_mode_0"), val = string("EXACT")]; + tensor input_191_cast_fp16 = gelu(mode = input_191_mode_0, x = input_189_cast_fp16)[name = string("input_191_cast_fp16")]; + string hidden_states_51_pad_type_0 = const()[name = string("hidden_states_51_pad_type_0"), val = string("valid")]; + tensor hidden_states_51_strides_0 = const()[name = string("hidden_states_51_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_51_pad_0 = const()[name = string("hidden_states_51_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_51_dilations_0 = const()[name = string("hidden_states_51_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_51_groups_0 = const()[name = string("hidden_states_51_groups_0"), val = int32(1)]; + tensor layers_23_fc2_weight_to_fp16 = const()[name = string("layers_23_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(946033024)))]; + tensor layers_23_fc2_bias_to_fp16 = const()[name = string("layers_23_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(959140288)))]; + tensor hidden_states_51_cast_fp16 = conv(bias = layers_23_fc2_bias_to_fp16, dilations = hidden_states_51_dilations_0, groups = hidden_states_51_groups_0, pad = hidden_states_51_pad_0, pad_type = hidden_states_51_pad_type_0, strides = hidden_states_51_strides_0, weight = layers_23_fc2_weight_to_fp16, x = input_191_cast_fp16)[name = string("hidden_states_51_cast_fp16")]; + tensor inputs_97_cast_fp16 = add(x = inputs_95_cast_fp16, y = hidden_states_51_cast_fp16)[name = string("inputs_97_cast_fp16")]; + int32 var_3014 = const()[name = string("op_3014"), val = int32(3)]; + tensor out_97_axes_0 = const()[name = string("out_97_axes_0"), val = tensor([1])]; + fp16 var_3033_to_fp16 = const()[name = string("op_3033_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_97_cast_fp16 = layer_norm(axes = out_97_axes_0, epsilon = var_3033_to_fp16, x = inputs_97_cast_fp16)[name = string("out_97_cast_fp16")]; + tensor obj_97_gamma_0_to_fp16 = const()[name = string("obj_97_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(959142912)))]; + tensor obj_97_beta_0_to_fp16 = const()[name = string("obj_97_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(959145536)))]; + fp16 obj_97_epsilon_0_to_fp16 = const()[name = string("obj_97_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor obj_97_cast_fp16 = batch_norm(beta = obj_97_beta_0_to_fp16, epsilon = obj_97_epsilon_0_to_fp16, gamma = obj_97_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_97_cast_fp16)[name = string("obj_97_cast_fp16")]; + string query_49_pad_type_0 = const()[name = string("query_49_pad_type_0"), val = string("valid")]; + tensor query_49_strides_0 = const()[name = string("query_49_strides_0"), val = tensor([1, 1])]; + tensor query_49_pad_0 = const()[name = string("query_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_49_dilations_0 = const()[name = string("query_49_dilations_0"), val = tensor([1, 1])]; + int32 query_49_groups_0 = const()[name = string("query_49_groups_0"), val = int32(1)]; + tensor layers_24_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_24_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(959148160)))]; + tensor layers_24_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_24_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(962425024)))]; + tensor query_49_cast_fp16 = conv(bias = layers_24_self_attn_q_proj_bias_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_24_self_attn_q_proj_weight_to_fp16, x = obj_97_cast_fp16)[name = string("query_49_cast_fp16")]; + string key_49_pad_type_0 = const()[name = string("key_49_pad_type_0"), val = string("valid")]; + tensor key_49_strides_0 = const()[name = string("key_49_strides_0"), val = tensor([1, 1])]; + tensor key_49_pad_0 = const()[name = string("key_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_49_dilations_0 = const()[name = string("key_49_dilations_0"), val = tensor([1, 1])]; + int32 key_49_groups_0 = const()[name = string("key_49_groups_0"), val = int32(1)]; + tensor layers_24_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_24_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(962427648)))]; + tensor key_49_cast_fp16 = conv(dilations = key_49_dilations_0, groups = key_49_groups_0, pad = key_49_pad_0, pad_type = key_49_pad_type_0, strides = key_49_strides_0, weight = layers_24_self_attn_k_proj_weight_to_fp16, x = obj_97_cast_fp16)[name = string("key_49_cast_fp16")]; + string value_49_pad_type_0 = const()[name = string("value_49_pad_type_0"), val = string("valid")]; + tensor value_49_strides_0 = const()[name = string("value_49_strides_0"), val = tensor([1, 1])]; + tensor value_49_pad_0 = const()[name = string("value_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_49_dilations_0 = const()[name = string("value_49_dilations_0"), val = tensor([1, 1])]; + int32 value_49_groups_0 = const()[name = string("value_49_groups_0"), val = int32(1)]; + tensor layers_24_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_24_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(965704512)))]; + tensor layers_24_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_24_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(968981376)))]; + tensor value_49_cast_fp16 = conv(bias = layers_24_self_attn_v_proj_bias_to_fp16, dilations = value_49_dilations_0, groups = value_49_groups_0, pad = value_49_pad_0, pad_type = value_49_pad_type_0, strides = value_49_strides_0, weight = layers_24_self_attn_v_proj_weight_to_fp16, x = obj_97_cast_fp16)[name = string("value_49_cast_fp16")]; + tensor var_3068 = const()[name = string("op_3068"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_49_cast_fp16 = reshape(shape = var_3068, x = query_49_cast_fp16)[name = string("mh_q_49_cast_fp16")]; + fp16 var_3070_to_fp16 = const()[name = string("op_3070_to_fp16"), val = fp16(0x1p-3)]; + tensor var_3071_cast_fp16 = mul(x = mh_q_49_cast_fp16, y = var_3070_to_fp16)[name = string("op_3071_cast_fp16")]; + tensor var_3072 = const()[name = string("op_3072"), val = tensor([1, 20, 64, -1])]; + tensor var_3073_cast_fp16 = reshape(shape = var_3072, x = key_49_cast_fp16)[name = string("op_3073_cast_fp16")]; + bool mh_w_49_transpose_x_0 = const()[name = string("mh_w_49_transpose_x_0"), val = bool(true)]; + bool mh_w_49_transpose_y_0 = const()[name = string("mh_w_49_transpose_y_0"), val = bool(false)]; + tensor mh_w_49_cast_fp16 = matmul(transpose_x = mh_w_49_transpose_x_0, transpose_y = mh_w_49_transpose_y_0, x = var_3071_cast_fp16, y = var_3073_cast_fp16)[name = string("mh_w_49_cast_fp16")]; + tensor var_3076_cast_fp16 = softmax(axis = var_3014, x = mh_w_49_cast_fp16)[name = string("op_3076_cast_fp16")]; + tensor var_3077 = const()[name = string("op_3077"), val = tensor([1, 20, 64, -1])]; + tensor var_3078_cast_fp16 = reshape(shape = var_3077, x = value_49_cast_fp16)[name = string("op_3078_cast_fp16")]; + bool attn_49_transpose_x_0 = const()[name = string("attn_49_transpose_x_0"), val = bool(false)]; + bool attn_49_transpose_y_0 = const()[name = string("attn_49_transpose_y_0"), val = bool(true)]; + tensor attn_49_cast_fp16 = matmul(transpose_x = attn_49_transpose_x_0, transpose_y = attn_49_transpose_y_0, x = var_3078_cast_fp16, y = var_3076_cast_fp16)[name = string("attn_49_cast_fp16")]; + tensor var_3081 = const()[name = string("op_3081"), val = tensor([1, 1280, 1, -1])]; + tensor input_193_cast_fp16 = reshape(shape = var_3081, x = attn_49_cast_fp16)[name = string("input_193_cast_fp16")]; + string obj_99_pad_type_0 = const()[name = string("obj_99_pad_type_0"), val = string("valid")]; + tensor obj_99_strides_0 = const()[name = string("obj_99_strides_0"), val = tensor([1, 1])]; + tensor obj_99_pad_0 = const()[name = string("obj_99_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_99_dilations_0 = const()[name = string("obj_99_dilations_0"), val = tensor([1, 1])]; + int32 obj_99_groups_0 = const()[name = string("obj_99_groups_0"), val = int32(1)]; + tensor layers_24_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_24_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(968984000)))]; + tensor layers_24_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_24_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(972260864)))]; + tensor obj_99_cast_fp16 = conv(bias = layers_24_self_attn_o_proj_bias_to_fp16, dilations = obj_99_dilations_0, groups = obj_99_groups_0, pad = obj_99_pad_0, pad_type = obj_99_pad_type_0, strides = obj_99_strides_0, weight = layers_24_self_attn_o_proj_weight_to_fp16, x = input_193_cast_fp16)[name = string("obj_99_cast_fp16")]; + tensor inputs_99_cast_fp16 = add(x = inputs_97_cast_fp16, y = obj_99_cast_fp16)[name = string("inputs_99_cast_fp16")]; + tensor out_99_axes_0 = const()[name = string("out_99_axes_0"), val = tensor([1])]; + fp16 var_3099_to_fp16 = const()[name = string("op_3099_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_99_cast_fp16 = layer_norm(axes = out_99_axes_0, epsilon = var_3099_to_fp16, x = inputs_99_cast_fp16)[name = string("out_99_cast_fp16")]; + tensor input_195_gamma_0_to_fp16 = const()[name = string("input_195_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(972263488)))]; + tensor input_195_beta_0_to_fp16 = const()[name = string("input_195_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(972266112)))]; + fp16 input_195_epsilon_0_to_fp16 = const()[name = string("input_195_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor input_195_cast_fp16 = batch_norm(beta = input_195_beta_0_to_fp16, epsilon = input_195_epsilon_0_to_fp16, gamma = input_195_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_99_cast_fp16)[name = string("input_195_cast_fp16")]; + string input_197_pad_type_0 = const()[name = string("input_197_pad_type_0"), val = string("valid")]; + tensor input_197_strides_0 = const()[name = string("input_197_strides_0"), val = tensor([1, 1])]; + tensor input_197_pad_0 = const()[name = string("input_197_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_197_dilations_0 = const()[name = string("input_197_dilations_0"), val = tensor([1, 1])]; + int32 input_197_groups_0 = const()[name = string("input_197_groups_0"), val = int32(1)]; + tensor layers_24_fc1_weight_to_fp16 = const()[name = string("layers_24_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(972268736)))]; + tensor layers_24_fc1_bias_to_fp16 = const()[name = string("layers_24_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(985376000)))]; + tensor input_197_cast_fp16 = conv(bias = layers_24_fc1_bias_to_fp16, dilations = input_197_dilations_0, groups = input_197_groups_0, pad = input_197_pad_0, pad_type = input_197_pad_type_0, strides = input_197_strides_0, weight = layers_24_fc1_weight_to_fp16, x = input_195_cast_fp16)[name = string("input_197_cast_fp16")]; + string input_199_mode_0 = const()[name = string("input_199_mode_0"), val = string("EXACT")]; + tensor input_199_cast_fp16 = gelu(mode = input_199_mode_0, x = input_197_cast_fp16)[name = string("input_199_cast_fp16")]; + string hidden_states_53_pad_type_0 = const()[name = string("hidden_states_53_pad_type_0"), val = string("valid")]; + tensor hidden_states_53_strides_0 = const()[name = string("hidden_states_53_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_53_pad_0 = const()[name = string("hidden_states_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_53_dilations_0 = const()[name = string("hidden_states_53_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_53_groups_0 = const()[name = string("hidden_states_53_groups_0"), val = int32(1)]; + tensor layers_24_fc2_weight_to_fp16 = const()[name = string("layers_24_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(985386304)))]; + tensor layers_24_fc2_bias_to_fp16 = const()[name = string("layers_24_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(998493568)))]; + tensor hidden_states_53_cast_fp16 = conv(bias = layers_24_fc2_bias_to_fp16, dilations = hidden_states_53_dilations_0, groups = hidden_states_53_groups_0, pad = hidden_states_53_pad_0, pad_type = hidden_states_53_pad_type_0, strides = hidden_states_53_strides_0, weight = layers_24_fc2_weight_to_fp16, x = input_199_cast_fp16)[name = string("hidden_states_53_cast_fp16")]; + tensor inputs_101_cast_fp16 = add(x = inputs_99_cast_fp16, y = hidden_states_53_cast_fp16)[name = string("inputs_101_cast_fp16")]; + int32 var_3132 = const()[name = string("op_3132"), val = int32(3)]; + tensor out_101_axes_0 = const()[name = string("out_101_axes_0"), val = tensor([1])]; + fp16 var_3151_to_fp16 = const()[name = string("op_3151_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_101_cast_fp16 = layer_norm(axes = out_101_axes_0, epsilon = var_3151_to_fp16, x = inputs_101_cast_fp16)[name = string("out_101_cast_fp16")]; + tensor obj_101_gamma_0_to_fp16 = const()[name = string("obj_101_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(998496192)))]; + tensor obj_101_beta_0_to_fp16 = const()[name = string("obj_101_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(998498816)))]; + fp16 obj_101_epsilon_0_to_fp16 = const()[name = string("obj_101_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor obj_101_cast_fp16 = batch_norm(beta = obj_101_beta_0_to_fp16, epsilon = obj_101_epsilon_0_to_fp16, gamma = obj_101_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_101_cast_fp16)[name = string("obj_101_cast_fp16")]; + string query_51_pad_type_0 = const()[name = string("query_51_pad_type_0"), val = string("valid")]; + tensor query_51_strides_0 = const()[name = string("query_51_strides_0"), val = tensor([1, 1])]; + tensor query_51_pad_0 = const()[name = string("query_51_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_51_dilations_0 = const()[name = string("query_51_dilations_0"), val = tensor([1, 1])]; + int32 query_51_groups_0 = const()[name = string("query_51_groups_0"), val = int32(1)]; + tensor layers_25_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_25_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(998501440)))]; + tensor layers_25_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_25_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1001778304)))]; + tensor query_51_cast_fp16 = conv(bias = layers_25_self_attn_q_proj_bias_to_fp16, dilations = query_51_dilations_0, groups = query_51_groups_0, pad = query_51_pad_0, pad_type = query_51_pad_type_0, strides = query_51_strides_0, weight = layers_25_self_attn_q_proj_weight_to_fp16, x = obj_101_cast_fp16)[name = string("query_51_cast_fp16")]; + string key_51_pad_type_0 = const()[name = string("key_51_pad_type_0"), val = string("valid")]; + tensor key_51_strides_0 = const()[name = string("key_51_strides_0"), val = tensor([1, 1])]; + tensor key_51_pad_0 = const()[name = string("key_51_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_51_dilations_0 = const()[name = string("key_51_dilations_0"), val = tensor([1, 1])]; + int32 key_51_groups_0 = const()[name = string("key_51_groups_0"), val = int32(1)]; + tensor layers_25_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_25_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1001780928)))]; + tensor key_51_cast_fp16 = conv(dilations = key_51_dilations_0, groups = key_51_groups_0, pad = key_51_pad_0, pad_type = key_51_pad_type_0, strides = key_51_strides_0, weight = layers_25_self_attn_k_proj_weight_to_fp16, x = obj_101_cast_fp16)[name = string("key_51_cast_fp16")]; + string value_51_pad_type_0 = const()[name = string("value_51_pad_type_0"), val = string("valid")]; + tensor value_51_strides_0 = const()[name = string("value_51_strides_0"), val = tensor([1, 1])]; + tensor value_51_pad_0 = const()[name = string("value_51_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_51_dilations_0 = const()[name = string("value_51_dilations_0"), val = tensor([1, 1])]; + int32 value_51_groups_0 = const()[name = string("value_51_groups_0"), val = int32(1)]; + tensor layers_25_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_25_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1005057792)))]; + tensor layers_25_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_25_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1008334656)))]; + tensor value_51_cast_fp16 = conv(bias = layers_25_self_attn_v_proj_bias_to_fp16, dilations = value_51_dilations_0, groups = value_51_groups_0, pad = value_51_pad_0, pad_type = value_51_pad_type_0, strides = value_51_strides_0, weight = layers_25_self_attn_v_proj_weight_to_fp16, x = obj_101_cast_fp16)[name = string("value_51_cast_fp16")]; + tensor var_3186 = const()[name = string("op_3186"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_51_cast_fp16 = reshape(shape = var_3186, x = query_51_cast_fp16)[name = string("mh_q_51_cast_fp16")]; + fp16 var_3188_to_fp16 = const()[name = string("op_3188_to_fp16"), val = fp16(0x1p-3)]; + tensor var_3189_cast_fp16 = mul(x = mh_q_51_cast_fp16, y = var_3188_to_fp16)[name = string("op_3189_cast_fp16")]; + tensor var_3190 = const()[name = string("op_3190"), val = tensor([1, 20, 64, -1])]; + tensor var_3191_cast_fp16 = reshape(shape = var_3190, x = key_51_cast_fp16)[name = string("op_3191_cast_fp16")]; + bool mh_w_51_transpose_x_0 = const()[name = string("mh_w_51_transpose_x_0"), val = bool(true)]; + bool mh_w_51_transpose_y_0 = const()[name = string("mh_w_51_transpose_y_0"), val = bool(false)]; + tensor mh_w_51_cast_fp16 = matmul(transpose_x = mh_w_51_transpose_x_0, transpose_y = mh_w_51_transpose_y_0, x = var_3189_cast_fp16, y = var_3191_cast_fp16)[name = string("mh_w_51_cast_fp16")]; + tensor var_3194_cast_fp16 = softmax(axis = var_3132, x = mh_w_51_cast_fp16)[name = string("op_3194_cast_fp16")]; + tensor var_3195 = const()[name = string("op_3195"), val = tensor([1, 20, 64, -1])]; + tensor var_3196_cast_fp16 = reshape(shape = var_3195, x = value_51_cast_fp16)[name = string("op_3196_cast_fp16")]; + bool attn_51_transpose_x_0 = const()[name = string("attn_51_transpose_x_0"), val = bool(false)]; + bool attn_51_transpose_y_0 = const()[name = string("attn_51_transpose_y_0"), val = bool(true)]; + tensor attn_51_cast_fp16 = matmul(transpose_x = attn_51_transpose_x_0, transpose_y = attn_51_transpose_y_0, x = var_3196_cast_fp16, y = var_3194_cast_fp16)[name = string("attn_51_cast_fp16")]; + tensor var_3199 = const()[name = string("op_3199"), val = tensor([1, 1280, 1, -1])]; + tensor input_201_cast_fp16 = reshape(shape = var_3199, x = attn_51_cast_fp16)[name = string("input_201_cast_fp16")]; + string obj_103_pad_type_0 = const()[name = string("obj_103_pad_type_0"), val = string("valid")]; + tensor obj_103_strides_0 = const()[name = string("obj_103_strides_0"), val = tensor([1, 1])]; + tensor obj_103_pad_0 = const()[name = string("obj_103_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_103_dilations_0 = const()[name = string("obj_103_dilations_0"), val = tensor([1, 1])]; + int32 obj_103_groups_0 = const()[name = string("obj_103_groups_0"), val = int32(1)]; + tensor layers_25_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_25_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1008337280)))]; + tensor layers_25_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_25_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1011614144)))]; + tensor obj_103_cast_fp16 = conv(bias = layers_25_self_attn_o_proj_bias_to_fp16, dilations = obj_103_dilations_0, groups = obj_103_groups_0, pad = obj_103_pad_0, pad_type = obj_103_pad_type_0, strides = obj_103_strides_0, weight = layers_25_self_attn_o_proj_weight_to_fp16, x = input_201_cast_fp16)[name = string("obj_103_cast_fp16")]; + tensor inputs_103_cast_fp16 = add(x = inputs_101_cast_fp16, y = obj_103_cast_fp16)[name = string("inputs_103_cast_fp16")]; + tensor out_103_axes_0 = const()[name = string("out_103_axes_0"), val = tensor([1])]; + fp16 var_3217_to_fp16 = const()[name = string("op_3217_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_103_cast_fp16 = layer_norm(axes = out_103_axes_0, epsilon = var_3217_to_fp16, x = inputs_103_cast_fp16)[name = string("out_103_cast_fp16")]; + tensor input_203_gamma_0_to_fp16 = const()[name = string("input_203_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1011616768)))]; + tensor input_203_beta_0_to_fp16 = const()[name = string("input_203_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1011619392)))]; + fp16 input_203_epsilon_0_to_fp16 = const()[name = string("input_203_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor input_203_cast_fp16 = batch_norm(beta = input_203_beta_0_to_fp16, epsilon = input_203_epsilon_0_to_fp16, gamma = input_203_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_103_cast_fp16)[name = string("input_203_cast_fp16")]; + string input_205_pad_type_0 = const()[name = string("input_205_pad_type_0"), val = string("valid")]; + tensor input_205_strides_0 = const()[name = string("input_205_strides_0"), val = tensor([1, 1])]; + tensor input_205_pad_0 = const()[name = string("input_205_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_205_dilations_0 = const()[name = string("input_205_dilations_0"), val = tensor([1, 1])]; + int32 input_205_groups_0 = const()[name = string("input_205_groups_0"), val = int32(1)]; + tensor layers_25_fc1_weight_to_fp16 = const()[name = string("layers_25_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1011622016)))]; + tensor layers_25_fc1_bias_to_fp16 = const()[name = string("layers_25_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1024729280)))]; + tensor input_205_cast_fp16 = conv(bias = layers_25_fc1_bias_to_fp16, dilations = input_205_dilations_0, groups = input_205_groups_0, pad = input_205_pad_0, pad_type = input_205_pad_type_0, strides = input_205_strides_0, weight = layers_25_fc1_weight_to_fp16, x = input_203_cast_fp16)[name = string("input_205_cast_fp16")]; + string input_207_mode_0 = const()[name = string("input_207_mode_0"), val = string("EXACT")]; + tensor input_207_cast_fp16 = gelu(mode = input_207_mode_0, x = input_205_cast_fp16)[name = string("input_207_cast_fp16")]; + string hidden_states_55_pad_type_0 = const()[name = string("hidden_states_55_pad_type_0"), val = string("valid")]; + tensor hidden_states_55_strides_0 = const()[name = string("hidden_states_55_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_55_pad_0 = const()[name = string("hidden_states_55_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_55_dilations_0 = const()[name = string("hidden_states_55_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_55_groups_0 = const()[name = string("hidden_states_55_groups_0"), val = int32(1)]; + tensor layers_25_fc2_weight_to_fp16 = const()[name = string("layers_25_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1024739584)))]; + tensor layers_25_fc2_bias_to_fp16 = const()[name = string("layers_25_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1037846848)))]; + tensor hidden_states_55_cast_fp16 = conv(bias = layers_25_fc2_bias_to_fp16, dilations = hidden_states_55_dilations_0, groups = hidden_states_55_groups_0, pad = hidden_states_55_pad_0, pad_type = hidden_states_55_pad_type_0, strides = hidden_states_55_strides_0, weight = layers_25_fc2_weight_to_fp16, x = input_207_cast_fp16)[name = string("hidden_states_55_cast_fp16")]; + tensor inputs_105_cast_fp16 = add(x = inputs_103_cast_fp16, y = hidden_states_55_cast_fp16)[name = string("inputs_105_cast_fp16")]; + int32 var_3250 = const()[name = string("op_3250"), val = int32(3)]; + tensor out_105_axes_0 = const()[name = string("out_105_axes_0"), val = tensor([1])]; + fp16 var_3269_to_fp16 = const()[name = string("op_3269_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_105_cast_fp16 = layer_norm(axes = out_105_axes_0, epsilon = var_3269_to_fp16, x = inputs_105_cast_fp16)[name = string("out_105_cast_fp16")]; + tensor obj_105_gamma_0_to_fp16 = const()[name = string("obj_105_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1037849472)))]; + tensor obj_105_beta_0_to_fp16 = const()[name = string("obj_105_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1037852096)))]; + fp16 obj_105_epsilon_0_to_fp16 = const()[name = string("obj_105_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor obj_105_cast_fp16 = batch_norm(beta = obj_105_beta_0_to_fp16, epsilon = obj_105_epsilon_0_to_fp16, gamma = obj_105_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_105_cast_fp16)[name = string("obj_105_cast_fp16")]; + string query_53_pad_type_0 = const()[name = string("query_53_pad_type_0"), val = string("valid")]; + tensor query_53_strides_0 = const()[name = string("query_53_strides_0"), val = tensor([1, 1])]; + tensor query_53_pad_0 = const()[name = string("query_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_53_dilations_0 = const()[name = string("query_53_dilations_0"), val = tensor([1, 1])]; + int32 query_53_groups_0 = const()[name = string("query_53_groups_0"), val = int32(1)]; + tensor layers_26_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_26_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1037854720)))]; + tensor layers_26_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_26_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1041131584)))]; + tensor query_53_cast_fp16 = conv(bias = layers_26_self_attn_q_proj_bias_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_26_self_attn_q_proj_weight_to_fp16, x = obj_105_cast_fp16)[name = string("query_53_cast_fp16")]; + string key_53_pad_type_0 = const()[name = string("key_53_pad_type_0"), val = string("valid")]; + tensor key_53_strides_0 = const()[name = string("key_53_strides_0"), val = tensor([1, 1])]; + tensor key_53_pad_0 = const()[name = string("key_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_53_dilations_0 = const()[name = string("key_53_dilations_0"), val = tensor([1, 1])]; + int32 key_53_groups_0 = const()[name = string("key_53_groups_0"), val = int32(1)]; + tensor layers_26_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_26_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1041134208)))]; + tensor key_53_cast_fp16 = conv(dilations = key_53_dilations_0, groups = key_53_groups_0, pad = key_53_pad_0, pad_type = key_53_pad_type_0, strides = key_53_strides_0, weight = layers_26_self_attn_k_proj_weight_to_fp16, x = obj_105_cast_fp16)[name = string("key_53_cast_fp16")]; + string value_53_pad_type_0 = const()[name = string("value_53_pad_type_0"), val = string("valid")]; + tensor value_53_strides_0 = const()[name = string("value_53_strides_0"), val = tensor([1, 1])]; + tensor value_53_pad_0 = const()[name = string("value_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_53_dilations_0 = const()[name = string("value_53_dilations_0"), val = tensor([1, 1])]; + int32 value_53_groups_0 = const()[name = string("value_53_groups_0"), val = int32(1)]; + tensor layers_26_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_26_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1044411072)))]; + tensor layers_26_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_26_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1047687936)))]; + tensor value_53_cast_fp16 = conv(bias = layers_26_self_attn_v_proj_bias_to_fp16, dilations = value_53_dilations_0, groups = value_53_groups_0, pad = value_53_pad_0, pad_type = value_53_pad_type_0, strides = value_53_strides_0, weight = layers_26_self_attn_v_proj_weight_to_fp16, x = obj_105_cast_fp16)[name = string("value_53_cast_fp16")]; + tensor var_3304 = const()[name = string("op_3304"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_53_cast_fp16 = reshape(shape = var_3304, x = query_53_cast_fp16)[name = string("mh_q_53_cast_fp16")]; + fp16 var_3306_to_fp16 = const()[name = string("op_3306_to_fp16"), val = fp16(0x1p-3)]; + tensor var_3307_cast_fp16 = mul(x = mh_q_53_cast_fp16, y = var_3306_to_fp16)[name = string("op_3307_cast_fp16")]; + tensor var_3308 = const()[name = string("op_3308"), val = tensor([1, 20, 64, -1])]; + tensor var_3309_cast_fp16 = reshape(shape = var_3308, x = key_53_cast_fp16)[name = string("op_3309_cast_fp16")]; + bool mh_w_53_transpose_x_0 = const()[name = string("mh_w_53_transpose_x_0"), val = bool(true)]; + bool mh_w_53_transpose_y_0 = const()[name = string("mh_w_53_transpose_y_0"), val = bool(false)]; + tensor mh_w_53_cast_fp16 = matmul(transpose_x = mh_w_53_transpose_x_0, transpose_y = mh_w_53_transpose_y_0, x = var_3307_cast_fp16, y = var_3309_cast_fp16)[name = string("mh_w_53_cast_fp16")]; + tensor var_3312_cast_fp16 = softmax(axis = var_3250, x = mh_w_53_cast_fp16)[name = string("op_3312_cast_fp16")]; + tensor var_3313 = const()[name = string("op_3313"), val = tensor([1, 20, 64, -1])]; + tensor var_3314_cast_fp16 = reshape(shape = var_3313, x = value_53_cast_fp16)[name = string("op_3314_cast_fp16")]; + bool attn_53_transpose_x_0 = const()[name = string("attn_53_transpose_x_0"), val = bool(false)]; + bool attn_53_transpose_y_0 = const()[name = string("attn_53_transpose_y_0"), val = bool(true)]; + tensor attn_53_cast_fp16 = matmul(transpose_x = attn_53_transpose_x_0, transpose_y = attn_53_transpose_y_0, x = var_3314_cast_fp16, y = var_3312_cast_fp16)[name = string("attn_53_cast_fp16")]; + tensor var_3317 = const()[name = string("op_3317"), val = tensor([1, 1280, 1, -1])]; + tensor input_209_cast_fp16 = reshape(shape = var_3317, x = attn_53_cast_fp16)[name = string("input_209_cast_fp16")]; + string obj_107_pad_type_0 = const()[name = string("obj_107_pad_type_0"), val = string("valid")]; + tensor obj_107_strides_0 = const()[name = string("obj_107_strides_0"), val = tensor([1, 1])]; + tensor obj_107_pad_0 = const()[name = string("obj_107_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_107_dilations_0 = const()[name = string("obj_107_dilations_0"), val = tensor([1, 1])]; + int32 obj_107_groups_0 = const()[name = string("obj_107_groups_0"), val = int32(1)]; + tensor layers_26_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_26_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1047690560)))]; + tensor layers_26_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_26_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1050967424)))]; + tensor obj_107_cast_fp16 = conv(bias = layers_26_self_attn_o_proj_bias_to_fp16, dilations = obj_107_dilations_0, groups = obj_107_groups_0, pad = obj_107_pad_0, pad_type = obj_107_pad_type_0, strides = obj_107_strides_0, weight = layers_26_self_attn_o_proj_weight_to_fp16, x = input_209_cast_fp16)[name = string("obj_107_cast_fp16")]; + tensor inputs_107_cast_fp16 = add(x = inputs_105_cast_fp16, y = obj_107_cast_fp16)[name = string("inputs_107_cast_fp16")]; + tensor out_107_axes_0 = const()[name = string("out_107_axes_0"), val = tensor([1])]; + fp16 var_3335_to_fp16 = const()[name = string("op_3335_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_107_cast_fp16 = layer_norm(axes = out_107_axes_0, epsilon = var_3335_to_fp16, x = inputs_107_cast_fp16)[name = string("out_107_cast_fp16")]; + tensor input_211_gamma_0_to_fp16 = const()[name = string("input_211_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1050970048)))]; + tensor input_211_beta_0_to_fp16 = const()[name = string("input_211_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1050972672)))]; + fp16 input_211_epsilon_0_to_fp16 = const()[name = string("input_211_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor input_211_cast_fp16 = batch_norm(beta = input_211_beta_0_to_fp16, epsilon = input_211_epsilon_0_to_fp16, gamma = input_211_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_107_cast_fp16)[name = string("input_211_cast_fp16")]; + string input_213_pad_type_0 = const()[name = string("input_213_pad_type_0"), val = string("valid")]; + tensor input_213_strides_0 = const()[name = string("input_213_strides_0"), val = tensor([1, 1])]; + tensor input_213_pad_0 = const()[name = string("input_213_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_213_dilations_0 = const()[name = string("input_213_dilations_0"), val = tensor([1, 1])]; + int32 input_213_groups_0 = const()[name = string("input_213_groups_0"), val = int32(1)]; + tensor layers_26_fc1_weight_to_fp16 = const()[name = string("layers_26_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1050975296)))]; + tensor layers_26_fc1_bias_to_fp16 = const()[name = string("layers_26_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1064082560)))]; + tensor input_213_cast_fp16 = conv(bias = layers_26_fc1_bias_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 = layers_26_fc1_weight_to_fp16, x = input_211_cast_fp16)[name = string("input_213_cast_fp16")]; + string input_215_mode_0 = const()[name = string("input_215_mode_0"), val = string("EXACT")]; + tensor input_215_cast_fp16 = gelu(mode = input_215_mode_0, x = input_213_cast_fp16)[name = string("input_215_cast_fp16")]; + string hidden_states_57_pad_type_0 = const()[name = string("hidden_states_57_pad_type_0"), val = string("valid")]; + tensor hidden_states_57_strides_0 = const()[name = string("hidden_states_57_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_57_pad_0 = const()[name = string("hidden_states_57_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_57_dilations_0 = const()[name = string("hidden_states_57_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_57_groups_0 = const()[name = string("hidden_states_57_groups_0"), val = int32(1)]; + tensor layers_26_fc2_weight_to_fp16 = const()[name = string("layers_26_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1064092864)))]; + tensor layers_26_fc2_bias_to_fp16 = const()[name = string("layers_26_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1077200128)))]; + tensor hidden_states_57_cast_fp16 = conv(bias = layers_26_fc2_bias_to_fp16, dilations = hidden_states_57_dilations_0, groups = hidden_states_57_groups_0, pad = hidden_states_57_pad_0, pad_type = hidden_states_57_pad_type_0, strides = hidden_states_57_strides_0, weight = layers_26_fc2_weight_to_fp16, x = input_215_cast_fp16)[name = string("hidden_states_57_cast_fp16")]; + tensor inputs_109_cast_fp16 = add(x = inputs_107_cast_fp16, y = hidden_states_57_cast_fp16)[name = string("inputs_109_cast_fp16")]; + int32 var_3368 = const()[name = string("op_3368"), val = int32(3)]; + tensor out_109_axes_0 = const()[name = string("out_109_axes_0"), val = tensor([1])]; + fp16 var_3387_to_fp16 = const()[name = string("op_3387_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_109_cast_fp16 = layer_norm(axes = out_109_axes_0, epsilon = var_3387_to_fp16, x = inputs_109_cast_fp16)[name = string("out_109_cast_fp16")]; + tensor obj_109_gamma_0_to_fp16 = const()[name = string("obj_109_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1077202752)))]; + tensor obj_109_beta_0_to_fp16 = const()[name = string("obj_109_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1077205376)))]; + fp16 obj_109_epsilon_0_to_fp16 = const()[name = string("obj_109_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor obj_109_cast_fp16 = batch_norm(beta = obj_109_beta_0_to_fp16, epsilon = obj_109_epsilon_0_to_fp16, gamma = obj_109_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_109_cast_fp16)[name = string("obj_109_cast_fp16")]; + string query_55_pad_type_0 = const()[name = string("query_55_pad_type_0"), val = string("valid")]; + tensor query_55_strides_0 = const()[name = string("query_55_strides_0"), val = tensor([1, 1])]; + tensor query_55_pad_0 = const()[name = string("query_55_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_55_dilations_0 = const()[name = string("query_55_dilations_0"), val = tensor([1, 1])]; + int32 query_55_groups_0 = const()[name = string("query_55_groups_0"), val = int32(1)]; + tensor layers_27_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_27_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1077208000)))]; + tensor layers_27_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_27_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1080484864)))]; + tensor query_55_cast_fp16 = conv(bias = layers_27_self_attn_q_proj_bias_to_fp16, dilations = query_55_dilations_0, groups = query_55_groups_0, pad = query_55_pad_0, pad_type = query_55_pad_type_0, strides = query_55_strides_0, weight = layers_27_self_attn_q_proj_weight_to_fp16, x = obj_109_cast_fp16)[name = string("query_55_cast_fp16")]; + string key_55_pad_type_0 = const()[name = string("key_55_pad_type_0"), val = string("valid")]; + tensor key_55_strides_0 = const()[name = string("key_55_strides_0"), val = tensor([1, 1])]; + tensor key_55_pad_0 = const()[name = string("key_55_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_55_dilations_0 = const()[name = string("key_55_dilations_0"), val = tensor([1, 1])]; + int32 key_55_groups_0 = const()[name = string("key_55_groups_0"), val = int32(1)]; + tensor layers_27_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_27_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1080487488)))]; + tensor key_55_cast_fp16 = conv(dilations = key_55_dilations_0, groups = key_55_groups_0, pad = key_55_pad_0, pad_type = key_55_pad_type_0, strides = key_55_strides_0, weight = layers_27_self_attn_k_proj_weight_to_fp16, x = obj_109_cast_fp16)[name = string("key_55_cast_fp16")]; + string value_55_pad_type_0 = const()[name = string("value_55_pad_type_0"), val = string("valid")]; + tensor value_55_strides_0 = const()[name = string("value_55_strides_0"), val = tensor([1, 1])]; + tensor value_55_pad_0 = const()[name = string("value_55_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_55_dilations_0 = const()[name = string("value_55_dilations_0"), val = tensor([1, 1])]; + int32 value_55_groups_0 = const()[name = string("value_55_groups_0"), val = int32(1)]; + tensor layers_27_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_27_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1083764352)))]; + tensor layers_27_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_27_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1087041216)))]; + tensor value_55_cast_fp16 = conv(bias = layers_27_self_attn_v_proj_bias_to_fp16, dilations = value_55_dilations_0, groups = value_55_groups_0, pad = value_55_pad_0, pad_type = value_55_pad_type_0, strides = value_55_strides_0, weight = layers_27_self_attn_v_proj_weight_to_fp16, x = obj_109_cast_fp16)[name = string("value_55_cast_fp16")]; + tensor var_3422 = const()[name = string("op_3422"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_55_cast_fp16 = reshape(shape = var_3422, x = query_55_cast_fp16)[name = string("mh_q_55_cast_fp16")]; + fp16 var_3424_to_fp16 = const()[name = string("op_3424_to_fp16"), val = fp16(0x1p-3)]; + tensor var_3425_cast_fp16 = mul(x = mh_q_55_cast_fp16, y = var_3424_to_fp16)[name = string("op_3425_cast_fp16")]; + tensor var_3426 = const()[name = string("op_3426"), val = tensor([1, 20, 64, -1])]; + tensor var_3427_cast_fp16 = reshape(shape = var_3426, x = key_55_cast_fp16)[name = string("op_3427_cast_fp16")]; + bool mh_w_55_transpose_x_0 = const()[name = string("mh_w_55_transpose_x_0"), val = bool(true)]; + bool mh_w_55_transpose_y_0 = const()[name = string("mh_w_55_transpose_y_0"), val = bool(false)]; + tensor mh_w_55_cast_fp16 = matmul(transpose_x = mh_w_55_transpose_x_0, transpose_y = mh_w_55_transpose_y_0, x = var_3425_cast_fp16, y = var_3427_cast_fp16)[name = string("mh_w_55_cast_fp16")]; + tensor var_3430_cast_fp16 = softmax(axis = var_3368, x = mh_w_55_cast_fp16)[name = string("op_3430_cast_fp16")]; + tensor var_3431 = const()[name = string("op_3431"), val = tensor([1, 20, 64, -1])]; + tensor var_3432_cast_fp16 = reshape(shape = var_3431, x = value_55_cast_fp16)[name = string("op_3432_cast_fp16")]; + bool attn_55_transpose_x_0 = const()[name = string("attn_55_transpose_x_0"), val = bool(false)]; + bool attn_55_transpose_y_0 = const()[name = string("attn_55_transpose_y_0"), val = bool(true)]; + tensor attn_55_cast_fp16 = matmul(transpose_x = attn_55_transpose_x_0, transpose_y = attn_55_transpose_y_0, x = var_3432_cast_fp16, y = var_3430_cast_fp16)[name = string("attn_55_cast_fp16")]; + tensor var_3435 = const()[name = string("op_3435"), val = tensor([1, 1280, 1, -1])]; + tensor input_217_cast_fp16 = reshape(shape = var_3435, x = attn_55_cast_fp16)[name = string("input_217_cast_fp16")]; + string obj_111_pad_type_0 = const()[name = string("obj_111_pad_type_0"), val = string("valid")]; + tensor obj_111_strides_0 = const()[name = string("obj_111_strides_0"), val = tensor([1, 1])]; + tensor obj_111_pad_0 = const()[name = string("obj_111_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_111_dilations_0 = const()[name = string("obj_111_dilations_0"), val = tensor([1, 1])]; + int32 obj_111_groups_0 = const()[name = string("obj_111_groups_0"), val = int32(1)]; + tensor layers_27_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_27_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1087043840)))]; + tensor layers_27_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_27_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1090320704)))]; + tensor obj_111_cast_fp16 = conv(bias = layers_27_self_attn_o_proj_bias_to_fp16, dilations = obj_111_dilations_0, groups = obj_111_groups_0, pad = obj_111_pad_0, pad_type = obj_111_pad_type_0, strides = obj_111_strides_0, weight = layers_27_self_attn_o_proj_weight_to_fp16, x = input_217_cast_fp16)[name = string("obj_111_cast_fp16")]; + tensor inputs_111_cast_fp16 = add(x = inputs_109_cast_fp16, y = obj_111_cast_fp16)[name = string("inputs_111_cast_fp16")]; + tensor out_111_axes_0 = const()[name = string("out_111_axes_0"), val = tensor([1])]; + fp16 var_3453_to_fp16 = const()[name = string("op_3453_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_111_cast_fp16 = layer_norm(axes = out_111_axes_0, epsilon = var_3453_to_fp16, x = inputs_111_cast_fp16)[name = string("out_111_cast_fp16")]; + tensor input_219_gamma_0_to_fp16 = const()[name = string("input_219_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1090323328)))]; + tensor input_219_beta_0_to_fp16 = const()[name = string("input_219_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1090325952)))]; + fp16 input_219_epsilon_0_to_fp16 = const()[name = string("input_219_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor 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 = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_111_cast_fp16)[name = string("input_219_cast_fp16")]; + string input_221_pad_type_0 = const()[name = string("input_221_pad_type_0"), val = string("valid")]; + tensor input_221_strides_0 = const()[name = string("input_221_strides_0"), val = tensor([1, 1])]; + tensor input_221_pad_0 = const()[name = string("input_221_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_221_dilations_0 = const()[name = string("input_221_dilations_0"), val = tensor([1, 1])]; + int32 input_221_groups_0 = const()[name = string("input_221_groups_0"), val = int32(1)]; + tensor layers_27_fc1_weight_to_fp16 = const()[name = string("layers_27_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1090328576)))]; + tensor layers_27_fc1_bias_to_fp16 = const()[name = string("layers_27_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1103435840)))]; + tensor input_221_cast_fp16 = conv(bias = layers_27_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_27_fc1_weight_to_fp16, x = input_219_cast_fp16)[name = string("input_221_cast_fp16")]; + string input_223_mode_0 = const()[name = string("input_223_mode_0"), val = string("EXACT")]; + tensor input_223_cast_fp16 = gelu(mode = input_223_mode_0, x = input_221_cast_fp16)[name = string("input_223_cast_fp16")]; + string hidden_states_59_pad_type_0 = const()[name = string("hidden_states_59_pad_type_0"), val = string("valid")]; + tensor hidden_states_59_strides_0 = const()[name = string("hidden_states_59_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_59_pad_0 = const()[name = string("hidden_states_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_59_dilations_0 = const()[name = string("hidden_states_59_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_59_groups_0 = const()[name = string("hidden_states_59_groups_0"), val = int32(1)]; + tensor layers_27_fc2_weight_to_fp16 = const()[name = string("layers_27_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1103446144)))]; + tensor layers_27_fc2_bias_to_fp16 = const()[name = string("layers_27_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1116553408)))]; + tensor hidden_states_59_cast_fp16 = conv(bias = layers_27_fc2_bias_to_fp16, dilations = hidden_states_59_dilations_0, groups = hidden_states_59_groups_0, pad = hidden_states_59_pad_0, pad_type = hidden_states_59_pad_type_0, strides = hidden_states_59_strides_0, weight = layers_27_fc2_weight_to_fp16, x = input_223_cast_fp16)[name = string("hidden_states_59_cast_fp16")]; + tensor inputs_113_cast_fp16 = add(x = inputs_111_cast_fp16, y = hidden_states_59_cast_fp16)[name = string("inputs_113_cast_fp16")]; + int32 var_3486 = const()[name = string("op_3486"), val = int32(3)]; + tensor out_113_axes_0 = const()[name = string("out_113_axes_0"), val = tensor([1])]; + fp16 var_3505_to_fp16 = const()[name = string("op_3505_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_113_cast_fp16 = layer_norm(axes = out_113_axes_0, epsilon = var_3505_to_fp16, x = inputs_113_cast_fp16)[name = string("out_113_cast_fp16")]; + tensor obj_113_gamma_0_to_fp16 = const()[name = string("obj_113_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1116556032)))]; + tensor obj_113_beta_0_to_fp16 = const()[name = string("obj_113_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1116558656)))]; + fp16 obj_113_epsilon_0_to_fp16 = const()[name = string("obj_113_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor obj_113_cast_fp16 = batch_norm(beta = obj_113_beta_0_to_fp16, epsilon = obj_113_epsilon_0_to_fp16, gamma = obj_113_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_113_cast_fp16)[name = string("obj_113_cast_fp16")]; + string query_57_pad_type_0 = const()[name = string("query_57_pad_type_0"), val = string("valid")]; + tensor query_57_strides_0 = const()[name = string("query_57_strides_0"), val = tensor([1, 1])]; + tensor query_57_pad_0 = const()[name = string("query_57_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_57_dilations_0 = const()[name = string("query_57_dilations_0"), val = tensor([1, 1])]; + int32 query_57_groups_0 = const()[name = string("query_57_groups_0"), val = int32(1)]; + tensor layers_28_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_28_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1116561280)))]; + tensor layers_28_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_28_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1119838144)))]; + tensor query_57_cast_fp16 = conv(bias = layers_28_self_attn_q_proj_bias_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_28_self_attn_q_proj_weight_to_fp16, x = obj_113_cast_fp16)[name = string("query_57_cast_fp16")]; + string key_57_pad_type_0 = const()[name = string("key_57_pad_type_0"), val = string("valid")]; + tensor key_57_strides_0 = const()[name = string("key_57_strides_0"), val = tensor([1, 1])]; + tensor key_57_pad_0 = const()[name = string("key_57_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_57_dilations_0 = const()[name = string("key_57_dilations_0"), val = tensor([1, 1])]; + int32 key_57_groups_0 = const()[name = string("key_57_groups_0"), val = int32(1)]; + tensor layers_28_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_28_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1119840768)))]; + tensor key_57_cast_fp16 = conv(dilations = key_57_dilations_0, groups = key_57_groups_0, pad = key_57_pad_0, pad_type = key_57_pad_type_0, strides = key_57_strides_0, weight = layers_28_self_attn_k_proj_weight_to_fp16, x = obj_113_cast_fp16)[name = string("key_57_cast_fp16")]; + string value_57_pad_type_0 = const()[name = string("value_57_pad_type_0"), val = string("valid")]; + tensor value_57_strides_0 = const()[name = string("value_57_strides_0"), val = tensor([1, 1])]; + tensor value_57_pad_0 = const()[name = string("value_57_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_57_dilations_0 = const()[name = string("value_57_dilations_0"), val = tensor([1, 1])]; + int32 value_57_groups_0 = const()[name = string("value_57_groups_0"), val = int32(1)]; + tensor layers_28_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_28_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1123117632)))]; + tensor layers_28_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_28_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1126394496)))]; + tensor value_57_cast_fp16 = conv(bias = layers_28_self_attn_v_proj_bias_to_fp16, dilations = value_57_dilations_0, groups = value_57_groups_0, pad = value_57_pad_0, pad_type = value_57_pad_type_0, strides = value_57_strides_0, weight = layers_28_self_attn_v_proj_weight_to_fp16, x = obj_113_cast_fp16)[name = string("value_57_cast_fp16")]; + tensor var_3540 = const()[name = string("op_3540"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_57_cast_fp16 = reshape(shape = var_3540, x = query_57_cast_fp16)[name = string("mh_q_57_cast_fp16")]; + fp16 var_3542_to_fp16 = const()[name = string("op_3542_to_fp16"), val = fp16(0x1p-3)]; + tensor var_3543_cast_fp16 = mul(x = mh_q_57_cast_fp16, y = var_3542_to_fp16)[name = string("op_3543_cast_fp16")]; + tensor var_3544 = const()[name = string("op_3544"), val = tensor([1, 20, 64, -1])]; + tensor var_3545_cast_fp16 = reshape(shape = var_3544, x = key_57_cast_fp16)[name = string("op_3545_cast_fp16")]; + bool mh_w_57_transpose_x_0 = const()[name = string("mh_w_57_transpose_x_0"), val = bool(true)]; + bool mh_w_57_transpose_y_0 = const()[name = string("mh_w_57_transpose_y_0"), val = bool(false)]; + tensor mh_w_57_cast_fp16 = matmul(transpose_x = mh_w_57_transpose_x_0, transpose_y = mh_w_57_transpose_y_0, x = var_3543_cast_fp16, y = var_3545_cast_fp16)[name = string("mh_w_57_cast_fp16")]; + tensor var_3548_cast_fp16 = softmax(axis = var_3486, x = mh_w_57_cast_fp16)[name = string("op_3548_cast_fp16")]; + tensor var_3549 = const()[name = string("op_3549"), val = tensor([1, 20, 64, -1])]; + tensor var_3550_cast_fp16 = reshape(shape = var_3549, x = value_57_cast_fp16)[name = string("op_3550_cast_fp16")]; + bool attn_57_transpose_x_0 = const()[name = string("attn_57_transpose_x_0"), val = bool(false)]; + bool attn_57_transpose_y_0 = const()[name = string("attn_57_transpose_y_0"), val = bool(true)]; + tensor attn_57_cast_fp16 = matmul(transpose_x = attn_57_transpose_x_0, transpose_y = attn_57_transpose_y_0, x = var_3550_cast_fp16, y = var_3548_cast_fp16)[name = string("attn_57_cast_fp16")]; + tensor var_3553 = const()[name = string("op_3553"), val = tensor([1, 1280, 1, -1])]; + tensor input_225_cast_fp16 = reshape(shape = var_3553, x = attn_57_cast_fp16)[name = string("input_225_cast_fp16")]; + string obj_115_pad_type_0 = const()[name = string("obj_115_pad_type_0"), val = string("valid")]; + tensor obj_115_strides_0 = const()[name = string("obj_115_strides_0"), val = tensor([1, 1])]; + tensor obj_115_pad_0 = const()[name = string("obj_115_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_115_dilations_0 = const()[name = string("obj_115_dilations_0"), val = tensor([1, 1])]; + int32 obj_115_groups_0 = const()[name = string("obj_115_groups_0"), val = int32(1)]; + tensor layers_28_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_28_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1126397120)))]; + tensor layers_28_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_28_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1129673984)))]; + tensor obj_115_cast_fp16 = conv(bias = layers_28_self_attn_o_proj_bias_to_fp16, dilations = obj_115_dilations_0, groups = obj_115_groups_0, pad = obj_115_pad_0, pad_type = obj_115_pad_type_0, strides = obj_115_strides_0, weight = layers_28_self_attn_o_proj_weight_to_fp16, x = input_225_cast_fp16)[name = string("obj_115_cast_fp16")]; + tensor inputs_115_cast_fp16 = add(x = inputs_113_cast_fp16, y = obj_115_cast_fp16)[name = string("inputs_115_cast_fp16")]; + tensor out_115_axes_0 = const()[name = string("out_115_axes_0"), val = tensor([1])]; + fp16 var_3571_to_fp16 = const()[name = string("op_3571_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_115_cast_fp16 = layer_norm(axes = out_115_axes_0, epsilon = var_3571_to_fp16, x = inputs_115_cast_fp16)[name = string("out_115_cast_fp16")]; + tensor input_227_gamma_0_to_fp16 = const()[name = string("input_227_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1129676608)))]; + tensor input_227_beta_0_to_fp16 = const()[name = string("input_227_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1129679232)))]; + fp16 input_227_epsilon_0_to_fp16 = const()[name = string("input_227_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor input_227_cast_fp16 = batch_norm(beta = input_227_beta_0_to_fp16, epsilon = input_227_epsilon_0_to_fp16, gamma = input_227_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_115_cast_fp16)[name = string("input_227_cast_fp16")]; + string input_229_pad_type_0 = const()[name = string("input_229_pad_type_0"), val = string("valid")]; + tensor input_229_strides_0 = const()[name = string("input_229_strides_0"), val = tensor([1, 1])]; + tensor input_229_pad_0 = const()[name = string("input_229_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_229_dilations_0 = const()[name = string("input_229_dilations_0"), val = tensor([1, 1])]; + int32 input_229_groups_0 = const()[name = string("input_229_groups_0"), val = int32(1)]; + tensor layers_28_fc1_weight_to_fp16 = const()[name = string("layers_28_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1129681856)))]; + tensor layers_28_fc1_bias_to_fp16 = const()[name = string("layers_28_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1142789120)))]; + tensor input_229_cast_fp16 = conv(bias = layers_28_fc1_bias_to_fp16, dilations = input_229_dilations_0, groups = input_229_groups_0, pad = input_229_pad_0, pad_type = input_229_pad_type_0, strides = input_229_strides_0, weight = layers_28_fc1_weight_to_fp16, x = input_227_cast_fp16)[name = string("input_229_cast_fp16")]; + string input_231_mode_0 = const()[name = string("input_231_mode_0"), val = string("EXACT")]; + tensor input_231_cast_fp16 = gelu(mode = input_231_mode_0, x = input_229_cast_fp16)[name = string("input_231_cast_fp16")]; + string hidden_states_61_pad_type_0 = const()[name = string("hidden_states_61_pad_type_0"), val = string("valid")]; + tensor hidden_states_61_strides_0 = const()[name = string("hidden_states_61_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_61_pad_0 = const()[name = string("hidden_states_61_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_61_dilations_0 = const()[name = string("hidden_states_61_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_61_groups_0 = const()[name = string("hidden_states_61_groups_0"), val = int32(1)]; + tensor layers_28_fc2_weight_to_fp16 = const()[name = string("layers_28_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1142799424)))]; + tensor layers_28_fc2_bias_to_fp16 = const()[name = string("layers_28_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1155906688)))]; + tensor hidden_states_61_cast_fp16 = conv(bias = layers_28_fc2_bias_to_fp16, dilations = hidden_states_61_dilations_0, groups = hidden_states_61_groups_0, pad = hidden_states_61_pad_0, pad_type = hidden_states_61_pad_type_0, strides = hidden_states_61_strides_0, weight = layers_28_fc2_weight_to_fp16, x = input_231_cast_fp16)[name = string("hidden_states_61_cast_fp16")]; + tensor inputs_117_cast_fp16 = add(x = inputs_115_cast_fp16, y = hidden_states_61_cast_fp16)[name = string("inputs_117_cast_fp16")]; + int32 var_3604 = const()[name = string("op_3604"), val = int32(3)]; + tensor out_117_axes_0 = const()[name = string("out_117_axes_0"), val = tensor([1])]; + fp16 var_3623_to_fp16 = const()[name = string("op_3623_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_117_cast_fp16 = layer_norm(axes = out_117_axes_0, epsilon = var_3623_to_fp16, x = inputs_117_cast_fp16)[name = string("out_117_cast_fp16")]; + tensor obj_117_gamma_0_to_fp16 = const()[name = string("obj_117_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1155909312)))]; + tensor obj_117_beta_0_to_fp16 = const()[name = string("obj_117_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1155911936)))]; + fp16 obj_117_epsilon_0_to_fp16 = const()[name = string("obj_117_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor obj_117_cast_fp16 = batch_norm(beta = obj_117_beta_0_to_fp16, epsilon = obj_117_epsilon_0_to_fp16, gamma = obj_117_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_117_cast_fp16)[name = string("obj_117_cast_fp16")]; + string query_59_pad_type_0 = const()[name = string("query_59_pad_type_0"), val = string("valid")]; + tensor query_59_strides_0 = const()[name = string("query_59_strides_0"), val = tensor([1, 1])]; + tensor query_59_pad_0 = const()[name = string("query_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_59_dilations_0 = const()[name = string("query_59_dilations_0"), val = tensor([1, 1])]; + int32 query_59_groups_0 = const()[name = string("query_59_groups_0"), val = int32(1)]; + tensor layers_29_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_29_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1155914560)))]; + tensor layers_29_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_29_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1159191424)))]; + tensor query_59_cast_fp16 = conv(bias = layers_29_self_attn_q_proj_bias_to_fp16, dilations = query_59_dilations_0, groups = query_59_groups_0, pad = query_59_pad_0, pad_type = query_59_pad_type_0, strides = query_59_strides_0, weight = layers_29_self_attn_q_proj_weight_to_fp16, x = obj_117_cast_fp16)[name = string("query_59_cast_fp16")]; + string key_59_pad_type_0 = const()[name = string("key_59_pad_type_0"), val = string("valid")]; + tensor key_59_strides_0 = const()[name = string("key_59_strides_0"), val = tensor([1, 1])]; + tensor key_59_pad_0 = const()[name = string("key_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_59_dilations_0 = const()[name = string("key_59_dilations_0"), val = tensor([1, 1])]; + int32 key_59_groups_0 = const()[name = string("key_59_groups_0"), val = int32(1)]; + tensor layers_29_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_29_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1159194048)))]; + tensor key_59_cast_fp16 = conv(dilations = key_59_dilations_0, groups = key_59_groups_0, pad = key_59_pad_0, pad_type = key_59_pad_type_0, strides = key_59_strides_0, weight = layers_29_self_attn_k_proj_weight_to_fp16, x = obj_117_cast_fp16)[name = string("key_59_cast_fp16")]; + string value_59_pad_type_0 = const()[name = string("value_59_pad_type_0"), val = string("valid")]; + tensor value_59_strides_0 = const()[name = string("value_59_strides_0"), val = tensor([1, 1])]; + tensor value_59_pad_0 = const()[name = string("value_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_59_dilations_0 = const()[name = string("value_59_dilations_0"), val = tensor([1, 1])]; + int32 value_59_groups_0 = const()[name = string("value_59_groups_0"), val = int32(1)]; + tensor layers_29_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_29_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1162470912)))]; + tensor layers_29_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_29_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1165747776)))]; + tensor value_59_cast_fp16 = conv(bias = layers_29_self_attn_v_proj_bias_to_fp16, dilations = value_59_dilations_0, groups = value_59_groups_0, pad = value_59_pad_0, pad_type = value_59_pad_type_0, strides = value_59_strides_0, weight = layers_29_self_attn_v_proj_weight_to_fp16, x = obj_117_cast_fp16)[name = string("value_59_cast_fp16")]; + tensor var_3658 = const()[name = string("op_3658"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_59_cast_fp16 = reshape(shape = var_3658, x = query_59_cast_fp16)[name = string("mh_q_59_cast_fp16")]; + fp16 var_3660_to_fp16 = const()[name = string("op_3660_to_fp16"), val = fp16(0x1p-3)]; + tensor var_3661_cast_fp16 = mul(x = mh_q_59_cast_fp16, y = var_3660_to_fp16)[name = string("op_3661_cast_fp16")]; + tensor var_3662 = const()[name = string("op_3662"), val = tensor([1, 20, 64, -1])]; + tensor var_3663_cast_fp16 = reshape(shape = var_3662, x = key_59_cast_fp16)[name = string("op_3663_cast_fp16")]; + bool mh_w_59_transpose_x_0 = const()[name = string("mh_w_59_transpose_x_0"), val = bool(true)]; + bool mh_w_59_transpose_y_0 = const()[name = string("mh_w_59_transpose_y_0"), val = bool(false)]; + tensor mh_w_59_cast_fp16 = matmul(transpose_x = mh_w_59_transpose_x_0, transpose_y = mh_w_59_transpose_y_0, x = var_3661_cast_fp16, y = var_3663_cast_fp16)[name = string("mh_w_59_cast_fp16")]; + tensor var_3666_cast_fp16 = softmax(axis = var_3604, x = mh_w_59_cast_fp16)[name = string("op_3666_cast_fp16")]; + tensor var_3667 = const()[name = string("op_3667"), val = tensor([1, 20, 64, -1])]; + tensor var_3668_cast_fp16 = reshape(shape = var_3667, x = value_59_cast_fp16)[name = string("op_3668_cast_fp16")]; + bool attn_59_transpose_x_0 = const()[name = string("attn_59_transpose_x_0"), val = bool(false)]; + bool attn_59_transpose_y_0 = const()[name = string("attn_59_transpose_y_0"), val = bool(true)]; + tensor attn_59_cast_fp16 = matmul(transpose_x = attn_59_transpose_x_0, transpose_y = attn_59_transpose_y_0, x = var_3668_cast_fp16, y = var_3666_cast_fp16)[name = string("attn_59_cast_fp16")]; + tensor var_3671 = const()[name = string("op_3671"), val = tensor([1, 1280, 1, -1])]; + tensor input_233_cast_fp16 = reshape(shape = var_3671, x = attn_59_cast_fp16)[name = string("input_233_cast_fp16")]; + string obj_119_pad_type_0 = const()[name = string("obj_119_pad_type_0"), val = string("valid")]; + tensor obj_119_strides_0 = const()[name = string("obj_119_strides_0"), val = tensor([1, 1])]; + tensor obj_119_pad_0 = const()[name = string("obj_119_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_119_dilations_0 = const()[name = string("obj_119_dilations_0"), val = tensor([1, 1])]; + int32 obj_119_groups_0 = const()[name = string("obj_119_groups_0"), val = int32(1)]; + tensor layers_29_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_29_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1165750400)))]; + tensor layers_29_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_29_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1169027264)))]; + tensor obj_119_cast_fp16 = conv(bias = layers_29_self_attn_o_proj_bias_to_fp16, dilations = obj_119_dilations_0, groups = obj_119_groups_0, pad = obj_119_pad_0, pad_type = obj_119_pad_type_0, strides = obj_119_strides_0, weight = layers_29_self_attn_o_proj_weight_to_fp16, x = input_233_cast_fp16)[name = string("obj_119_cast_fp16")]; + tensor inputs_119_cast_fp16 = add(x = inputs_117_cast_fp16, y = obj_119_cast_fp16)[name = string("inputs_119_cast_fp16")]; + tensor out_119_axes_0 = const()[name = string("out_119_axes_0"), val = tensor([1])]; + fp16 var_3689_to_fp16 = const()[name = string("op_3689_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_119_cast_fp16 = layer_norm(axes = out_119_axes_0, epsilon = var_3689_to_fp16, x = inputs_119_cast_fp16)[name = string("out_119_cast_fp16")]; + tensor input_235_gamma_0_to_fp16 = const()[name = string("input_235_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1169029888)))]; + tensor input_235_beta_0_to_fp16 = const()[name = string("input_235_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1169032512)))]; + fp16 input_235_epsilon_0_to_fp16 = const()[name = string("input_235_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor input_235_cast_fp16 = batch_norm(beta = input_235_beta_0_to_fp16, epsilon = input_235_epsilon_0_to_fp16, gamma = input_235_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_119_cast_fp16)[name = string("input_235_cast_fp16")]; + string input_237_pad_type_0 = const()[name = string("input_237_pad_type_0"), val = string("valid")]; + tensor input_237_strides_0 = const()[name = string("input_237_strides_0"), val = tensor([1, 1])]; + tensor input_237_pad_0 = const()[name = string("input_237_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_237_dilations_0 = const()[name = string("input_237_dilations_0"), val = tensor([1, 1])]; + int32 input_237_groups_0 = const()[name = string("input_237_groups_0"), val = int32(1)]; + tensor layers_29_fc1_weight_to_fp16 = const()[name = string("layers_29_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1169035136)))]; + tensor layers_29_fc1_bias_to_fp16 = const()[name = string("layers_29_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1182142400)))]; + tensor input_237_cast_fp16 = conv(bias = layers_29_fc1_bias_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 = layers_29_fc1_weight_to_fp16, x = input_235_cast_fp16)[name = string("input_237_cast_fp16")]; + string input_239_mode_0 = const()[name = string("input_239_mode_0"), val = string("EXACT")]; + tensor input_239_cast_fp16 = gelu(mode = input_239_mode_0, x = input_237_cast_fp16)[name = string("input_239_cast_fp16")]; + string hidden_states_63_pad_type_0 = const()[name = string("hidden_states_63_pad_type_0"), val = string("valid")]; + tensor hidden_states_63_strides_0 = const()[name = string("hidden_states_63_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_63_pad_0 = const()[name = string("hidden_states_63_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_63_dilations_0 = const()[name = string("hidden_states_63_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_63_groups_0 = const()[name = string("hidden_states_63_groups_0"), val = int32(1)]; + tensor layers_29_fc2_weight_to_fp16 = const()[name = string("layers_29_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1182152704)))]; + tensor layers_29_fc2_bias_to_fp16 = const()[name = string("layers_29_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1195259968)))]; + tensor hidden_states_63_cast_fp16 = conv(bias = layers_29_fc2_bias_to_fp16, dilations = hidden_states_63_dilations_0, groups = hidden_states_63_groups_0, pad = hidden_states_63_pad_0, pad_type = hidden_states_63_pad_type_0, strides = hidden_states_63_strides_0, weight = layers_29_fc2_weight_to_fp16, x = input_239_cast_fp16)[name = string("hidden_states_63_cast_fp16")]; + tensor inputs_121_cast_fp16 = add(x = inputs_119_cast_fp16, y = hidden_states_63_cast_fp16)[name = string("inputs_121_cast_fp16")]; + int32 var_3722 = const()[name = string("op_3722"), val = int32(3)]; + tensor out_121_axes_0 = const()[name = string("out_121_axes_0"), val = tensor([1])]; + fp16 var_3741_to_fp16 = const()[name = string("op_3741_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_121_cast_fp16 = layer_norm(axes = out_121_axes_0, epsilon = var_3741_to_fp16, x = inputs_121_cast_fp16)[name = string("out_121_cast_fp16")]; + tensor obj_121_gamma_0_to_fp16 = const()[name = string("obj_121_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1195262592)))]; + tensor obj_121_beta_0_to_fp16 = const()[name = string("obj_121_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1195265216)))]; + fp16 obj_121_epsilon_0_to_fp16 = const()[name = string("obj_121_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor obj_121_cast_fp16 = batch_norm(beta = obj_121_beta_0_to_fp16, epsilon = obj_121_epsilon_0_to_fp16, gamma = obj_121_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_121_cast_fp16)[name = string("obj_121_cast_fp16")]; + string query_61_pad_type_0 = const()[name = string("query_61_pad_type_0"), val = string("valid")]; + tensor query_61_strides_0 = const()[name = string("query_61_strides_0"), val = tensor([1, 1])]; + tensor query_61_pad_0 = const()[name = string("query_61_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_61_dilations_0 = const()[name = string("query_61_dilations_0"), val = tensor([1, 1])]; + int32 query_61_groups_0 = const()[name = string("query_61_groups_0"), val = int32(1)]; + tensor layers_30_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_30_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1195267840)))]; + tensor layers_30_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_30_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1198544704)))]; + tensor query_61_cast_fp16 = conv(bias = layers_30_self_attn_q_proj_bias_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_30_self_attn_q_proj_weight_to_fp16, x = obj_121_cast_fp16)[name = string("query_61_cast_fp16")]; + string key_61_pad_type_0 = const()[name = string("key_61_pad_type_0"), val = string("valid")]; + tensor key_61_strides_0 = const()[name = string("key_61_strides_0"), val = tensor([1, 1])]; + tensor key_61_pad_0 = const()[name = string("key_61_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_61_dilations_0 = const()[name = string("key_61_dilations_0"), val = tensor([1, 1])]; + int32 key_61_groups_0 = const()[name = string("key_61_groups_0"), val = int32(1)]; + tensor layers_30_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_30_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1198547328)))]; + tensor key_61_cast_fp16 = conv(dilations = key_61_dilations_0, groups = key_61_groups_0, pad = key_61_pad_0, pad_type = key_61_pad_type_0, strides = key_61_strides_0, weight = layers_30_self_attn_k_proj_weight_to_fp16, x = obj_121_cast_fp16)[name = string("key_61_cast_fp16")]; + string value_61_pad_type_0 = const()[name = string("value_61_pad_type_0"), val = string("valid")]; + tensor value_61_strides_0 = const()[name = string("value_61_strides_0"), val = tensor([1, 1])]; + tensor value_61_pad_0 = const()[name = string("value_61_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_61_dilations_0 = const()[name = string("value_61_dilations_0"), val = tensor([1, 1])]; + int32 value_61_groups_0 = const()[name = string("value_61_groups_0"), val = int32(1)]; + tensor layers_30_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_30_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1201824192)))]; + tensor layers_30_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_30_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1205101056)))]; + tensor value_61_cast_fp16 = conv(bias = layers_30_self_attn_v_proj_bias_to_fp16, dilations = value_61_dilations_0, groups = value_61_groups_0, pad = value_61_pad_0, pad_type = value_61_pad_type_0, strides = value_61_strides_0, weight = layers_30_self_attn_v_proj_weight_to_fp16, x = obj_121_cast_fp16)[name = string("value_61_cast_fp16")]; + tensor var_3776 = const()[name = string("op_3776"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_61_cast_fp16 = reshape(shape = var_3776, x = query_61_cast_fp16)[name = string("mh_q_61_cast_fp16")]; + fp16 var_3778_to_fp16 = const()[name = string("op_3778_to_fp16"), val = fp16(0x1p-3)]; + tensor var_3779_cast_fp16 = mul(x = mh_q_61_cast_fp16, y = var_3778_to_fp16)[name = string("op_3779_cast_fp16")]; + tensor var_3780 = const()[name = string("op_3780"), val = tensor([1, 20, 64, -1])]; + tensor var_3781_cast_fp16 = reshape(shape = var_3780, x = key_61_cast_fp16)[name = string("op_3781_cast_fp16")]; + bool mh_w_61_transpose_x_0 = const()[name = string("mh_w_61_transpose_x_0"), val = bool(true)]; + bool mh_w_61_transpose_y_0 = const()[name = string("mh_w_61_transpose_y_0"), val = bool(false)]; + tensor mh_w_61_cast_fp16 = matmul(transpose_x = mh_w_61_transpose_x_0, transpose_y = mh_w_61_transpose_y_0, x = var_3779_cast_fp16, y = var_3781_cast_fp16)[name = string("mh_w_61_cast_fp16")]; + tensor var_3784_cast_fp16 = softmax(axis = var_3722, x = mh_w_61_cast_fp16)[name = string("op_3784_cast_fp16")]; + tensor var_3785 = const()[name = string("op_3785"), val = tensor([1, 20, 64, -1])]; + tensor var_3786_cast_fp16 = reshape(shape = var_3785, x = value_61_cast_fp16)[name = string("op_3786_cast_fp16")]; + bool attn_61_transpose_x_0 = const()[name = string("attn_61_transpose_x_0"), val = bool(false)]; + bool attn_61_transpose_y_0 = const()[name = string("attn_61_transpose_y_0"), val = bool(true)]; + tensor attn_61_cast_fp16 = matmul(transpose_x = attn_61_transpose_x_0, transpose_y = attn_61_transpose_y_0, x = var_3786_cast_fp16, y = var_3784_cast_fp16)[name = string("attn_61_cast_fp16")]; + tensor var_3789 = const()[name = string("op_3789"), val = tensor([1, 1280, 1, -1])]; + tensor input_241_cast_fp16 = reshape(shape = var_3789, x = attn_61_cast_fp16)[name = string("input_241_cast_fp16")]; + string obj_123_pad_type_0 = const()[name = string("obj_123_pad_type_0"), val = string("valid")]; + tensor obj_123_strides_0 = const()[name = string("obj_123_strides_0"), val = tensor([1, 1])]; + tensor obj_123_pad_0 = const()[name = string("obj_123_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_123_dilations_0 = const()[name = string("obj_123_dilations_0"), val = tensor([1, 1])]; + int32 obj_123_groups_0 = const()[name = string("obj_123_groups_0"), val = int32(1)]; + tensor layers_30_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_30_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1205103680)))]; + tensor layers_30_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_30_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1208380544)))]; + tensor obj_123_cast_fp16 = conv(bias = layers_30_self_attn_o_proj_bias_to_fp16, dilations = obj_123_dilations_0, groups = obj_123_groups_0, pad = obj_123_pad_0, pad_type = obj_123_pad_type_0, strides = obj_123_strides_0, weight = layers_30_self_attn_o_proj_weight_to_fp16, x = input_241_cast_fp16)[name = string("obj_123_cast_fp16")]; + tensor inputs_123_cast_fp16 = add(x = inputs_121_cast_fp16, y = obj_123_cast_fp16)[name = string("inputs_123_cast_fp16")]; + tensor out_123_axes_0 = const()[name = string("out_123_axes_0"), val = tensor([1])]; + fp16 var_3807_to_fp16 = const()[name = string("op_3807_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_123_cast_fp16 = layer_norm(axes = out_123_axes_0, epsilon = var_3807_to_fp16, x = inputs_123_cast_fp16)[name = string("out_123_cast_fp16")]; + tensor input_243_gamma_0_to_fp16 = const()[name = string("input_243_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1208383168)))]; + tensor input_243_beta_0_to_fp16 = const()[name = string("input_243_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1208385792)))]; + fp16 input_243_epsilon_0_to_fp16 = const()[name = string("input_243_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor 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 = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_123_cast_fp16)[name = string("input_243_cast_fp16")]; + string input_245_pad_type_0 = const()[name = string("input_245_pad_type_0"), val = string("valid")]; + tensor input_245_strides_0 = const()[name = string("input_245_strides_0"), val = tensor([1, 1])]; + tensor input_245_pad_0 = const()[name = string("input_245_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_245_dilations_0 = const()[name = string("input_245_dilations_0"), val = tensor([1, 1])]; + int32 input_245_groups_0 = const()[name = string("input_245_groups_0"), val = int32(1)]; + tensor layers_30_fc1_weight_to_fp16 = const()[name = string("layers_30_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1208388416)))]; + tensor layers_30_fc1_bias_to_fp16 = const()[name = string("layers_30_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1221495680)))]; + tensor input_245_cast_fp16 = conv(bias = layers_30_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_30_fc1_weight_to_fp16, x = input_243_cast_fp16)[name = string("input_245_cast_fp16")]; + string input_247_mode_0 = const()[name = string("input_247_mode_0"), val = string("EXACT")]; + tensor input_247_cast_fp16 = gelu(mode = input_247_mode_0, x = input_245_cast_fp16)[name = string("input_247_cast_fp16")]; + string hidden_states_65_pad_type_0 = const()[name = string("hidden_states_65_pad_type_0"), val = string("valid")]; + tensor hidden_states_65_strides_0 = const()[name = string("hidden_states_65_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_65_pad_0 = const()[name = string("hidden_states_65_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_65_dilations_0 = const()[name = string("hidden_states_65_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_65_groups_0 = const()[name = string("hidden_states_65_groups_0"), val = int32(1)]; + tensor layers_30_fc2_weight_to_fp16 = const()[name = string("layers_30_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1221505984)))]; + tensor layers_30_fc2_bias_to_fp16 = const()[name = string("layers_30_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1234613248)))]; + tensor hidden_states_65_cast_fp16 = conv(bias = layers_30_fc2_bias_to_fp16, dilations = hidden_states_65_dilations_0, groups = hidden_states_65_groups_0, pad = hidden_states_65_pad_0, pad_type = hidden_states_65_pad_type_0, strides = hidden_states_65_strides_0, weight = layers_30_fc2_weight_to_fp16, x = input_247_cast_fp16)[name = string("hidden_states_65_cast_fp16")]; + tensor inputs_125_cast_fp16 = add(x = inputs_123_cast_fp16, y = hidden_states_65_cast_fp16)[name = string("inputs_125_cast_fp16")]; + int32 var_3840 = const()[name = string("op_3840"), val = int32(3)]; + tensor out_125_axes_0 = const()[name = string("out_125_axes_0"), val = tensor([1])]; + fp16 var_3859_to_fp16 = const()[name = string("op_3859_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_125_cast_fp16 = layer_norm(axes = out_125_axes_0, epsilon = var_3859_to_fp16, x = inputs_125_cast_fp16)[name = string("out_125_cast_fp16")]; + tensor obj_125_gamma_0_to_fp16 = const()[name = string("obj_125_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1234615872)))]; + tensor obj_125_beta_0_to_fp16 = const()[name = string("obj_125_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1234618496)))]; + fp16 obj_125_epsilon_0_to_fp16 = const()[name = string("obj_125_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor obj_125_cast_fp16 = batch_norm(beta = obj_125_beta_0_to_fp16, epsilon = obj_125_epsilon_0_to_fp16, gamma = obj_125_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_125_cast_fp16)[name = string("obj_125_cast_fp16")]; + string query_pad_type_0 = const()[name = string("query_pad_type_0"), val = string("valid")]; + tensor query_strides_0 = const()[name = string("query_strides_0"), val = tensor([1, 1])]; + tensor query_pad_0 = const()[name = string("query_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_dilations_0 = const()[name = string("query_dilations_0"), val = tensor([1, 1])]; + int32 query_groups_0 = const()[name = string("query_groups_0"), val = int32(1)]; + tensor layers_31_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_31_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1234621120)))]; + tensor layers_31_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_31_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1237897984)))]; + tensor query_cast_fp16 = conv(bias = layers_31_self_attn_q_proj_bias_to_fp16, dilations = query_dilations_0, groups = query_groups_0, pad = query_pad_0, pad_type = query_pad_type_0, strides = query_strides_0, weight = layers_31_self_attn_q_proj_weight_to_fp16, x = obj_125_cast_fp16)[name = string("query_cast_fp16")]; + string key_pad_type_0 = const()[name = string("key_pad_type_0"), val = string("valid")]; + tensor key_strides_0 = const()[name = string("key_strides_0"), val = tensor([1, 1])]; + tensor key_pad_0 = const()[name = string("key_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_dilations_0 = const()[name = string("key_dilations_0"), val = tensor([1, 1])]; + int32 key_groups_0 = const()[name = string("key_groups_0"), val = int32(1)]; + tensor layers_31_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_31_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1237900608)))]; + tensor key_cast_fp16 = conv(dilations = key_dilations_0, groups = key_groups_0, pad = key_pad_0, pad_type = key_pad_type_0, strides = key_strides_0, weight = layers_31_self_attn_k_proj_weight_to_fp16, x = obj_125_cast_fp16)[name = string("key_cast_fp16")]; + string value_pad_type_0 = const()[name = string("value_pad_type_0"), val = string("valid")]; + tensor value_strides_0 = const()[name = string("value_strides_0"), val = tensor([1, 1])]; + tensor value_pad_0 = const()[name = string("value_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_dilations_0 = const()[name = string("value_dilations_0"), val = tensor([1, 1])]; + int32 value_groups_0 = const()[name = string("value_groups_0"), val = int32(1)]; + tensor layers_31_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_31_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1241177472)))]; + tensor layers_31_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_31_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1244454336)))]; + tensor value_cast_fp16 = conv(bias = layers_31_self_attn_v_proj_bias_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_31_self_attn_v_proj_weight_to_fp16, x = obj_125_cast_fp16)[name = string("value_cast_fp16")]; + tensor var_3894 = const()[name = string("op_3894"), val = tensor([1, 20, 64, -1])]; + tensor mh_q_cast_fp16 = reshape(shape = var_3894, x = query_cast_fp16)[name = string("mh_q_cast_fp16")]; + fp16 var_3896_to_fp16 = const()[name = string("op_3896_to_fp16"), val = fp16(0x1p-3)]; + tensor var_3897_cast_fp16 = mul(x = mh_q_cast_fp16, y = var_3896_to_fp16)[name = string("op_3897_cast_fp16")]; + tensor var_3898 = const()[name = string("op_3898"), val = tensor([1, 20, 64, -1])]; + tensor var_3899_cast_fp16 = reshape(shape = var_3898, x = key_cast_fp16)[name = string("op_3899_cast_fp16")]; + bool mh_w_transpose_x_0 = const()[name = string("mh_w_transpose_x_0"), val = bool(true)]; + bool mh_w_transpose_y_0 = const()[name = string("mh_w_transpose_y_0"), val = bool(false)]; + tensor mh_w_cast_fp16 = matmul(transpose_x = mh_w_transpose_x_0, transpose_y = mh_w_transpose_y_0, x = var_3897_cast_fp16, y = var_3899_cast_fp16)[name = string("mh_w_cast_fp16")]; + tensor var_3902_cast_fp16 = softmax(axis = var_3840, x = mh_w_cast_fp16)[name = string("op_3902_cast_fp16")]; + tensor var_3903 = const()[name = string("op_3903"), val = tensor([1, 20, 64, -1])]; + tensor var_3904_cast_fp16 = reshape(shape = var_3903, x = value_cast_fp16)[name = string("op_3904_cast_fp16")]; + bool attn_transpose_x_0 = const()[name = string("attn_transpose_x_0"), val = bool(false)]; + bool attn_transpose_y_0 = const()[name = string("attn_transpose_y_0"), val = bool(true)]; + tensor attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = var_3904_cast_fp16, y = var_3902_cast_fp16)[name = string("attn_cast_fp16")]; + tensor var_3907 = const()[name = string("op_3907"), val = tensor([1, 1280, 1, -1])]; + tensor input_249_cast_fp16 = reshape(shape = var_3907, x = attn_cast_fp16)[name = string("input_249_cast_fp16")]; + string obj_pad_type_0 = const()[name = string("obj_pad_type_0"), val = string("valid")]; + tensor obj_strides_0 = const()[name = string("obj_strides_0"), val = tensor([1, 1])]; + tensor obj_pad_0 = const()[name = string("obj_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_dilations_0 = const()[name = string("obj_dilations_0"), val = tensor([1, 1])]; + int32 obj_groups_0 = const()[name = string("obj_groups_0"), val = int32(1)]; + tensor layers_31_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_31_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1244456960)))]; + tensor layers_31_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_31_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1247733824)))]; + tensor obj_cast_fp16 = conv(bias = layers_31_self_attn_o_proj_bias_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_31_self_attn_o_proj_weight_to_fp16, x = input_249_cast_fp16)[name = string("obj_cast_fp16")]; + tensor inputs_127_cast_fp16 = add(x = inputs_125_cast_fp16, y = obj_cast_fp16)[name = string("inputs_127_cast_fp16")]; + tensor out_127_axes_0 = const()[name = string("out_127_axes_0"), val = tensor([1])]; + fp16 var_3925_to_fp16 = const()[name = string("op_3925_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_127_cast_fp16 = layer_norm(axes = out_127_axes_0, epsilon = var_3925_to_fp16, x = inputs_127_cast_fp16)[name = string("out_127_cast_fp16")]; + tensor input_251_gamma_0_to_fp16 = const()[name = string("input_251_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1247736448)))]; + tensor input_251_beta_0_to_fp16 = const()[name = string("input_251_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1247739072)))]; + fp16 input_251_epsilon_0_to_fp16 = const()[name = string("input_251_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor 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 = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_127_cast_fp16)[name = string("input_251_cast_fp16")]; + string input_253_pad_type_0 = const()[name = string("input_253_pad_type_0"), val = string("valid")]; + tensor input_253_strides_0 = const()[name = string("input_253_strides_0"), val = tensor([1, 1])]; + tensor input_253_pad_0 = const()[name = string("input_253_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_253_dilations_0 = const()[name = string("input_253_dilations_0"), val = tensor([1, 1])]; + int32 input_253_groups_0 = const()[name = string("input_253_groups_0"), val = int32(1)]; + tensor layers_31_fc1_weight_to_fp16 = const()[name = string("layers_31_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1247741696)))]; + tensor layers_31_fc1_bias_to_fp16 = const()[name = string("layers_31_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1260848960)))]; + tensor input_253_cast_fp16 = conv(bias = layers_31_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_31_fc1_weight_to_fp16, x = input_251_cast_fp16)[name = string("input_253_cast_fp16")]; + string input_255_mode_0 = const()[name = string("input_255_mode_0"), val = string("EXACT")]; + tensor input_255_cast_fp16 = gelu(mode = input_255_mode_0, x = input_253_cast_fp16)[name = string("input_255_cast_fp16")]; + string hidden_states_pad_type_0 = const()[name = string("hidden_states_pad_type_0"), val = string("valid")]; + tensor hidden_states_strides_0 = const()[name = string("hidden_states_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_pad_0 = const()[name = string("hidden_states_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_dilations_0 = const()[name = string("hidden_states_dilations_0"), val = tensor([1, 1])]; + int32 hidden_states_groups_0 = const()[name = string("hidden_states_groups_0"), val = int32(1)]; + tensor layers_31_fc2_weight_to_fp16 = const()[name = string("layers_31_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1260859264)))]; + tensor layers_31_fc2_bias_to_fp16 = const()[name = string("layers_31_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1273966528)))]; + tensor hidden_states_cast_fp16 = conv(bias = layers_31_fc2_bias_to_fp16, dilations = hidden_states_dilations_0, groups = hidden_states_groups_0, pad = hidden_states_pad_0, pad_type = hidden_states_pad_type_0, strides = hidden_states_strides_0, weight = layers_31_fc2_weight_to_fp16, x = input_255_cast_fp16)[name = string("hidden_states_cast_fp16")]; + tensor inputs_cast_fp16 = add(x = inputs_127_cast_fp16, y = hidden_states_cast_fp16)[name = string("inputs_cast_fp16")]; + tensor out_axes_0 = const()[name = string("out_axes_0"), val = tensor([1])]; + fp16 var_3963_to_fp16 = const()[name = string("op_3963_to_fp16"), val = fp16(0x1.5p-17)]; + tensor out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_3963_to_fp16, x = inputs_cast_fp16)[name = string("out_cast_fp16")]; + tensor encoder_output_embeds_type_fp32_gamma_0_to_fp16 = const()[name = string("encoder_output_embeds_type_fp32_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1273969152)))]; + tensor encoder_output_embeds_type_fp32_beta_0_to_fp16 = const()[name = string("encoder_output_embeds_type_fp32_beta_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1273971776)))]; + fp16 encoder_output_embeds_type_fp32_epsilon_0_to_fp16 = const()[name = string("encoder_output_embeds_type_fp32_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)]; + tensor 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 = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_cast_fp16)[name = string("encoder_output_embeds_type_fp32_cast_fp16")]; + string var_3987_pad_type_0 = const()[name = string("op_3987_pad_type_0"), val = string("valid")]; + tensor var_3987_strides_0 = const()[name = string("op_3987_strides_0"), val = tensor([1, 1])]; + tensor var_3987_pad_0 = const()[name = string("op_3987_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3987_dilations_0 = const()[name = string("op_3987_dilations_0"), val = tensor([1, 1])]; + int32 var_3987_groups_0 = const()[name = string("op_3987_groups_0"), val = int32(1)]; + tensor decoder_kv_cache_prep_0_encoder_attn_k_proj_weight_to_fp16 = const()[name = string("decoder_kv_cache_prep_0_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1273974400)))]; + tensor var_3987_cast_fp16 = conv(dilations = var_3987_dilations_0, groups = var_3987_groups_0, pad = var_3987_pad_0, pad_type = var_3987_pad_type_0, strides = var_3987_strides_0, weight = decoder_kv_cache_prep_0_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = string("op_3987_cast_fp16")]; + string var_3994_pad_type_0 = const()[name = string("op_3994_pad_type_0"), val = string("valid")]; + tensor var_3994_strides_0 = const()[name = string("op_3994_strides_0"), val = tensor([1, 1])]; + tensor var_3994_pad_0 = const()[name = string("op_3994_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3994_dilations_0 = const()[name = string("op_3994_dilations_0"), val = tensor([1, 1])]; + int32 var_3994_groups_0 = const()[name = string("op_3994_groups_0"), val = int32(1)]; + tensor decoder_kv_cache_prep_0_encoder_attn_v_proj_weight_to_fp16 = const()[name = string("decoder_kv_cache_prep_0_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1277251264)))]; + tensor decoder_kv_cache_prep_0_encoder_attn_v_proj_bias_to_fp16 = const()[name = string("decoder_kv_cache_prep_0_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1280528128)))]; + tensor var_3994_cast_fp16 = conv(bias = decoder_kv_cache_prep_0_encoder_attn_v_proj_bias_to_fp16, dilations = var_3994_dilations_0, groups = var_3994_groups_0, pad = var_3994_pad_0, pad_type = var_3994_pad_type_0, strides = var_3994_strides_0, weight = decoder_kv_cache_prep_0_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = string("op_3994_cast_fp16")]; + string k_pad_type_0 = const()[name = string("k_pad_type_0"), val = string("valid")]; + tensor k_strides_0 = const()[name = string("k_strides_0"), val = tensor([1, 1])]; + tensor k_pad_0 = const()[name = string("k_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_dilations_0 = const()[name = string("k_dilations_0"), val = tensor([1, 1])]; + int32 k_groups_0 = const()[name = string("k_groups_0"), val = int32(1)]; + tensor decoder_kv_cache_prep_1_encoder_attn_k_proj_weight_to_fp16 = const()[name = string("decoder_kv_cache_prep_1_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1280530752)))]; + tensor k_cast_fp16 = conv(dilations = k_dilations_0, groups = k_groups_0, pad = k_pad_0, pad_type = k_pad_type_0, strides = k_strides_0, weight = decoder_kv_cache_prep_1_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = string("k_cast_fp16")]; + string v_pad_type_0 = const()[name = string("v_pad_type_0"), val = string("valid")]; + tensor v_strides_0 = const()[name = string("v_strides_0"), val = tensor([1, 1])]; + tensor v_pad_0 = const()[name = string("v_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor v_dilations_0 = const()[name = string("v_dilations_0"), val = tensor([1, 1])]; + int32 v_groups_0 = const()[name = string("v_groups_0"), val = int32(1)]; + tensor decoder_kv_cache_prep_1_encoder_attn_v_proj_weight_to_fp16 = const()[name = string("decoder_kv_cache_prep_1_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1283807616)))]; + tensor decoder_kv_cache_prep_1_encoder_attn_v_proj_bias_to_fp16 = const()[name = string("decoder_kv_cache_prep_1_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1287084480)))]; + tensor v_cast_fp16 = conv(bias = decoder_kv_cache_prep_1_encoder_attn_v_proj_bias_to_fp16, dilations = v_dilations_0, groups = v_groups_0, pad = v_pad_0, pad_type = v_pad_type_0, strides = v_strides_0, weight = decoder_kv_cache_prep_1_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = string("v_cast_fp16")]; + int32 var_4024 = const()[name = string("op_4024"), val = int32(0)]; + bool input_259_interleave_0 = const()[name = string("input_259_interleave_0"), val = bool(false)]; + tensor input_259_cast_fp16 = concat(axis = var_4024, interleave = input_259_interleave_0, values = (var_3987_cast_fp16, k_cast_fp16))[name = string("input_259_cast_fp16")]; + int32 var_4027 = const()[name = string("op_4027"), val = int32(0)]; + bool input_interleave_0 = const()[name = string("input_interleave_0"), val = bool(false)]; + tensor input_cast_fp16 = concat(axis = var_4027, interleave = input_interleave_0, values = (var_3994_cast_fp16, v_cast_fp16))[name = string("input_cast_fp16")]; + tensor var_4034_pad_0 = const()[name = string("op_4034_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 0, 36])]; + string var_4034_mode_0 = const()[name = string("op_4034_mode_0"), val = string("constant")]; + fp16 const_33_to_fp16 = const()[name = string("const_33_to_fp16"), val = fp16(0x0p+0)]; + tensor encoder_attn_key_cache = pad(constant_val = const_33_to_fp16, mode = var_4034_mode_0, pad = var_4034_pad_0, x = input_259_cast_fp16)[name = string("op_4034_cast_fp16")]; + tensor var_4040_pad_0 = const()[name = string("op_4040_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 0, 36])]; + string var_4040_mode_0 = const()[name = string("op_4040_mode_0"), val = string("constant")]; + fp16 const_34_to_fp16 = const()[name = string("const_34_to_fp16"), val = fp16(0x0p+0)]; + tensor encoder_attn_value_cache = pad(constant_val = const_34_to_fp16, mode = var_4040_mode_0, pad = var_4040_pad_0, x = input_cast_fp16)[name = string("op_4040_cast_fp16")]; + } -> (encoder_output_embeds, encoder_attn_key_cache, encoder_attn_value_cache); +} \ No newline at end of file