Synchronizing local compiler cache.

.gitattributes

@@ -3438,3 +3438,31 @@ neuronxcc-2.19.8089.0+8ab9f450/MODULE_e752b6d369dd13e651a9+ed72d204/model.neff f
 neuronxcc-2.19.8089.0+8ab9f450/MODULE_5ba89bbb07ae57870f9d+ed72d204/model.neff filter=lfs diff=lfs merge=lfs -text
 neuronxcc-2.19.8089.0+8ab9f450/MODULE_7b3c7ebc2d0673f07a89+a9d440f5/model.neff filter=lfs diff=lfs merge=lfs -text
 neuronxcc-2.19.8089.0+8ab9f450/MODULE_7b3c7ebc2d0673f07a89+a9d440f5/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.17.194.0+d312836f/MODULE_0319ecab9682f42a2e0f+bfe5714b/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.17.194.0+d312836f/MODULE_24e323acf78ef5eda141+165e9558/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.17.194.0+d312836f/MODULE_24e323acf78ef5eda141+165e9558/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.17.194.0+d312836f/MODULE_313796881a2973827e21+431f5505/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.17.194.0+d312836f/MODULE_4242784510489555269+e30acd3a/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.17.194.0+d312836f/MODULE_6708f635794756c6254f+165e9558/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.17.194.0+d312836f/MODULE_6708f635794756c6254f+165e9558/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.17.194.0+d312836f/MODULE_71a7e6b24554e80fa037+165e9558/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.17.194.0+d312836f/MODULE_71a7e6b24554e80fa037+165e9558/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.17.194.0+d312836f/MODULE_7ff50e368ec1bcc50af3+165e9558/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.17.194.0+d312836f/MODULE_7ff50e368ec1bcc50af3+165e9558/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.17.194.0+d312836f/MODULE_8704cbaf0ac16e5d3ac0+bfe5714b/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.17.194.0+d312836f/MODULE_89f903d621d68d3a2cf4+165e9558/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.17.194.0+d312836f/MODULE_89f903d621d68d3a2cf4+165e9558/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.17.194.0+d312836f/MODULE_975d14574546c97dc994+165e9558/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.17.194.0+d312836f/MODULE_975d14574546c97dc994+165e9558/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.17.194.0+d312836f/MODULE_9b17218ed71a6f1040ba+165e9558/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.17.194.0+d312836f/MODULE_9b17218ed71a6f1040ba+165e9558/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.17.194.0+d312836f/MODULE_9edc34cdab9401830e43+165e9558/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.17.194.0+d312836f/MODULE_9edc34cdab9401830e43+165e9558/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.17.194.0+d312836f/MODULE_a9fd63d01e95973c2777+165e9558/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.17.194.0+d312836f/MODULE_a9fd63d01e95973c2777+165e9558/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.17.194.0+d312836f/MODULE_b267bb985f633378c63a+165e9558/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.17.194.0+d312836f/MODULE_b267bb985f633378c63a+165e9558/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.17.194.0+d312836f/MODULE_cc7a21f76bfb2321fb54+165e9558/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.17.194.0+d312836f/MODULE_cc7a21f76bfb2321fb54+165e9558/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.17.194.0+d312836f/MODULE_cca573e7f7cd333a1619+165e9558/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.17.194.0+d312836f/MODULE_cca573e7f7cd333a1619+165e9558/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text

neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev2/llama/Maykeye/TinyLLama-v0/d8f06ba7bd2d5076813a.json

@@ -0,0 +1,71 @@
+{
+  "_entry_class": "SingleModelCacheEntry",
+  "_model_id": "Maykeye/TinyLLama-v0",
+  "_task": "text-generation",
+  "architectures": [
+    "LlamaForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "head_dim": 4,
+  "hidden_act": "silu",
+  "hidden_size": 64,
+  "initializer_range": 0.02,
+  "intermediate_size": 256,
+  "max_position_embeddings": 2048,
+  "mlp_bias": false,
+  "model_type": "llama",
+  "neuron": {
+    "_serialized_key": "NxDNeuronConfig",
+    "async_mode": false,
+    "attn_kernel_enabled": false,
+    "batch_size": 1,
+    "capacity_factor": null,
+    "cc_pipeline_tiling_factor": 2,
+    "checkpoint_id": "Maykeye/TinyLLama-v0",
+    "checkpoint_revision": "298338802ab94432b917bcce11382aa151aee50f",
+    "continuous_batching": false,
+    "enable_bucketing": false,
+    "ep_degree": 1,
+    "flash_decoding_enabled": false,
+    "fused_qkv": true,
+    "glu_mlp": true,
+    "is_chunked_prefill": false,
+    "local_ranks_size": 1,
+    "logical_nc_config": 1,
+    "max_batch_size": 1,
+    "max_context_length": 100,
+    "max_topk": 256,
+    "mlp_kernel_enabled": false,
+    "mlp_kernel_fuse_residual_add": false,
+    "n_active_tokens": 100,
+    "neuronxcc_version": "2.17.194.0+d312836f",
+    "num_cores_per_group": 1,
+    "on_device_sampling": true,
+    "optimum_neuron_version": "0.3.0.dev2",
+    "output_logits": false,
+    "padding_side": "right",
+    "pp_degree": 1,
+    "qk_layernorm": false,
+    "qkv_kernel_enabled": false,
+    "rpl_reduce_dtype": "bfloat16",
+    "sequence_length": 100,
+    "sequence_parallel_enabled": false,
+    "speculation_length": 0,
+    "start_rank_id": 0,
+    "target": null,
+    "torch_dtype": "bfloat16",
+    "tp_degree": 1,
+    "vocab_parallel": false
+  },
+  "num_attention_heads": 16,
+  "num_hidden_layers": 8,
+  "num_key_value_heads": 16,
+  "pretraining_tp": 1,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 10000.0,
+  "tie_word_embeddings": false,
+  "use_cache": true,
+  "vocab_size": 32000
+}

neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev2/llama/Maykeye/TinyLLama-v0/f1af079d52e1e2ac6a68.json

@@ -0,0 +1,71 @@
+{
+  "_entry_class": "SingleModelCacheEntry",
+  "_model_id": "Maykeye/TinyLLama-v0",
+  "_task": "text-generation",
+  "architectures": [
+    "LlamaForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "head_dim": 4,
+  "hidden_act": "silu",
+  "hidden_size": 64,
+  "initializer_range": 0.02,
+  "intermediate_size": 256,
+  "max_position_embeddings": 2048,
+  "mlp_bias": false,
+  "model_type": "llama",
+  "neuron": {
+    "_serialized_key": "NxDNeuronConfig",
+    "async_mode": false,
+    "attn_kernel_enabled": false,
+    "batch_size": 1,
+    "capacity_factor": null,
+    "cc_pipeline_tiling_factor": 2,
+    "checkpoint_id": "Maykeye/TinyLLama-v0",
+    "checkpoint_revision": "298338802ab94432b917bcce11382aa151aee50f",
+    "continuous_batching": false,
+    "enable_bucketing": false,
+    "ep_degree": 1,
+    "flash_decoding_enabled": false,
+    "fused_qkv": true,
+    "glu_mlp": true,
+    "is_chunked_prefill": false,
+    "local_ranks_size": 2,
+    "logical_nc_config": 1,
+    "max_batch_size": 1,
+    "max_context_length": 100,
+    "max_topk": 256,
+    "mlp_kernel_enabled": false,
+    "mlp_kernel_fuse_residual_add": false,
+    "n_active_tokens": 100,
+    "neuronxcc_version": "2.17.194.0+d312836f",
+    "num_cores_per_group": 1,
+    "on_device_sampling": true,
+    "optimum_neuron_version": "0.3.0.dev2",
+    "output_logits": false,
+    "padding_side": "right",
+    "pp_degree": 1,
+    "qk_layernorm": false,
+    "qkv_kernel_enabled": false,
+    "rpl_reduce_dtype": "bfloat16",
+    "sequence_length": 100,
+    "sequence_parallel_enabled": false,
+    "speculation_length": 0,
+    "start_rank_id": 0,
+    "target": null,
+    "torch_dtype": "bfloat16",
+    "tp_degree": 2,
+    "vocab_parallel": false
+  },
+  "num_attention_heads": 16,
+  "num_hidden_layers": 8,
+  "num_key_value_heads": 16,
+  "pretraining_tp": 1,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 10000.0,
+  "tie_word_embeddings": false,
+  "use_cache": true,
+  "vocab_size": 32000
+}

neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.3.0.dev4/granite/ibm-granite/granite-3.1-2b-instruct/c7927a430db3da5413ec.json

@@ -0,0 +1,73 @@
+{
+  "_entry_class": "SingleModelCacheEntry",
+  "_model_id": "ibm-granite/granite-3.1-2b-instruct",
+  "_task": "text-generation",
+  "architectures": [
+    "GraniteForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.1,
+  "attention_multiplier": 0.015625,
+  "embedding_multiplier": 12.0,
+  "hidden_act": "silu",
+  "hidden_size": 2048,
+  "initializer_range": 0.02,
+  "intermediate_size": 8192,
+  "logits_scaling": 8.0,
+  "max_position_embeddings": 131072,
+  "mlp_bias": false,
+  "model_type": "granite",
+  "neuron": {
+    "_serialized_key": "NxDNeuronConfig",
+    "async_mode": false,
+    "attn_kernel_enabled": false,
+    "batch_size": 4,
+    "capacity_factor": null,
+    "cc_pipeline_tiling_factor": 2,
+    "checkpoint_id": "ibm-granite/granite-3.1-2b-instruct",
+    "checkpoint_revision": "bbc2aed595bd38bd770263dc3ab831db9794441d",
+    "continuous_batching": true,
+    "enable_bucketing": false,
+    "ep_degree": 1,
+    "flash_decoding_enabled": false,
+    "fused_qkv": true,
+    "glu_mlp": true,
+    "is_chunked_prefill": false,
+    "local_ranks_size": 2,
+    "logical_nc_config": 1,
+    "max_batch_size": 4,
+    "max_context_length": 4096,
+    "max_topk": 256,
+    "mlp_kernel_enabled": false,
+    "mlp_kernel_fuse_residual_add": false,
+    "n_active_tokens": 4096,
+    "neuronxcc_version": "2.17.194.0+d312836f",
+    "num_cores_per_group": 1,
+    "on_device_sampling": false,
+    "optimum_neuron_version": "0.3.0.dev4",
+    "output_logits": false,
+    "padding_side": "right",
+    "pp_degree": 1,
+    "qk_layernorm": false,
+    "qkv_kernel_enabled": false,
+    "rpl_reduce_dtype": "bfloat16",
+    "sequence_length": 4096,
+    "sequence_parallel_enabled": false,
+    "speculation_length": 0,
+    "start_rank_id": 0,
+    "target": null,
+    "torch_dtype": "bfloat16",
+    "tp_degree": 2,
+    "vocab_parallel": false
+  },
+  "num_attention_heads": 32,
+  "num_hidden_layers": 40,
+  "num_key_value_heads": 8,
+  "residual_multiplier": 0.22,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": null,
+  "rope_theta": 5000000.0,
+  "tie_word_embeddings": true,
+  "use_cache": true,
+  "vocab_size": 49155
+}

neuronxcc-2.17.194.0+d312836f/MODULE_0319ecab9682f42a2e0f+bfe5714b/compile_flags.json

@@ -0,0 +1 @@
+"--auto-cast=none --model-type=transformer --tensorizer-options='--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-dge-dma --vectorize-strided-dma ' -O2  --internal-num-neuroncores-per-sengine=1 --logfile=/tmp/nxd_model/context_encoding_model/_tp0_bk0/log-neuron-cc.txt"

neuronxcc-2.17.194.0+d312836f/MODULE_0319ecab9682f42a2e0f+bfe5714b/model.hlo_module.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:934c016f62b55da98d34e488bae8668f0b80b8d8cffca713f1b6fdfb9fccd7da
+size 172596

neuronxcc-2.17.194.0+d312836f/MODULE_0319ecab9682f42a2e0f+bfe5714b/model.neff

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2983205a24f11d46e328ab17234974421fc4f4e5a8101d18d37978d42f27ae24
+size 246784

neuronxcc-2.17.194.0+d312836f/MODULE_12627483106468160534+e30acd3a/compile_flags.json

@@ -0,0 +1 @@
+["--target=trn1"]

neuronxcc-2.17.194.0+d312836f/MODULE_12627483106468160534+e30acd3a/model.hlo_module.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:002e7281f5713839b44010c4b90bfe6476fd6e72e142bfc561d0de77bb7c5347
+size 444

neuronxcc-2.17.194.0+d312836f/MODULE_14de102a7c791bd95a24+bfe5714b/compile_flags.json

@@ -0,0 +1 @@
+"--auto-cast=none --model-type=transformer --tensorizer-options='--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-dge-dma --vectorize-strided-dma ' -O2  --internal-num-neuroncores-per-sengine=1 --logfile=/tmp/nxd_model/context_encoding_model/_tp0_bk0/log-neuron-cc.txt"

neuronxcc-2.17.194.0+d312836f/MODULE_14de102a7c791bd95a24+bfe5714b/model.hlo_module.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:14ae716665e1a9c8c9d6031d935826ba4851b717286599b7a4c58d954408bd73
+size 18983459

neuronxcc-2.17.194.0+d312836f/MODULE_24e323acf78ef5eda141+165e9558/compile_flags.json

@@ -0,0 +1 @@
+"--auto-cast=none --model-type=transformer --tensorizer-options='--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-dge-dma --vectorize-strided-dma ' -O2  --internal-num-neuroncores-per-sengine=1 --logfile=/tmp/nxd_model/token_generation_model/_tp0_bk0/log-neuron-cc.txt --enable-internal-neff-wrapper"

neuronxcc-2.17.194.0+d312836f/MODULE_24e323acf78ef5eda141+165e9558/model.hlo_module.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:128a9a7fa2e0978562d62144e988bae68a25f5540399826d15122d09e8bcefad
+size 94870

neuronxcc-2.17.194.0+d312836f/MODULE_24e323acf78ef5eda141+165e9558/model.neff

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4688b6fdcd8bb37955378d8d814f6308a576b67b1fbeb99ad9f01f03af54364b
+size 697344

neuronxcc-2.17.194.0+d312836f/MODULE_24e323acf78ef5eda141+165e9558/wrapped_neff.hlo

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a004b787431d850901f0d55e2f907a20bd1ae0641492e86aaa2de713d41168c4
+size 719418

neuronxcc-2.17.194.0+d312836f/MODULE_2598f2596057e1346614+bfe5714b/compile_flags.json

@@ -0,0 +1 @@
+"--auto-cast=none --model-type=transformer --tensorizer-options='--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-dge-dma --vectorize-strided-dma ' -O2  --internal-num-neuroncores-per-sengine=1 --logfile=/tmp/nxd_model/context_encoding_model/_tp0_bk0/log-neuron-cc.txt"

neuronxcc-2.17.194.0+d312836f/MODULE_2598f2596057e1346614+bfe5714b/model.hlo_module.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1f0e5b28f8af837880b36cc46c7841e32fa2d3d182bd446f42aa4c64c12dc984
+size 52564523

neuronxcc-2.17.194.0+d312836f/MODULE_29474e39f37931275586+431f5505/compile_flags.json

@@ -0,0 +1 @@
+"--model-type=transformer -O1 --lnc=1 --internal-hlo2tensorizer-options=--experimental-unsafe-fp8e4m3fn-as-fp8e4m3 --logfile=/tmp/nxd_model/layout_opt/log-neuron-cc.txt"

neuronxcc-2.17.194.0+d312836f/MODULE_29474e39f37931275586+431f5505/model.hlo_module.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ad46fd05ebf2e4017087e0f3d35d9fddbb73ab8d311d8b92a3aab43cfc49aaf4
+size 25857

neuronxcc-2.17.194.0+d312836f/MODULE_313796881a2973827e21+431f5505/compile_flags.json

@@ -0,0 +1 @@
+"--model-type=transformer -O1 --lnc=1 --internal-hlo2tensorizer-options=--experimental-unsafe-fp8e4m3fn-as-fp8e4m3 --logfile=/tmp/nxd_model/layout_opt/log-neuron-cc.txt"

neuronxcc-2.17.194.0+d312836f/MODULE_313796881a2973827e21+431f5505/model.hlo_module.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a8f1e698fa23c2e5951644290686c8f13b4beb75a9227b943568f8cdebad7796
+size 172361

neuronxcc-2.17.194.0+d312836f/MODULE_313796881a2973827e21+431f5505/model.neff

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8e02708ac46a0c44c69b51532b77058ff134f18ca27729d890d80af8f0212e2e
+size 2243584

neuronxcc-2.17.194.0+d312836f/MODULE_4242784510489555269+e30acd3a/compile_flags.json

@@ -0,0 +1 @@
+["--target=trn1"]

neuronxcc-2.17.194.0+d312836f/MODULE_4242784510489555269+e30acd3a/model.hlo_module.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c96868686ca6df16e3a140bd0cf614740d8817d40ea268707bdbb4724ac5a7cf
+size 75607

neuronxcc-2.17.194.0+d312836f/MODULE_4242784510489555269+e30acd3a/model.neff

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:20a07c8d2be4d29dcf50a9b108f0d3f8360ac63be1248806704fa075c1cfc551
+size 7097344

neuronxcc-2.17.194.0+d312836f/MODULE_4585214586228082616+e30acd3a/compile_flags.json

@@ -0,0 +1 @@
+["--target=trn1"]

neuronxcc-2.17.194.0+d312836f/MODULE_4585214586228082616+e30acd3a/model.hlo_module.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b432b2f785ef8e9c680909d44b1f4a240b786a1921ca0694a010f5cc8e34a33a
+size 488

neuronxcc-2.17.194.0+d312836f/MODULE_49b8463ac655cfcc0b60+165e9558/compile_flags.json

@@ -0,0 +1 @@
+"--auto-cast=none --model-type=transformer --tensorizer-options='--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-dge-dma --vectorize-strided-dma ' -O2  --internal-num-neuroncores-per-sengine=1 --logfile=/tmp/nxd_model/token_generation_model/_tp0_bk0/log-neuron-cc.txt --enable-internal-neff-wrapper"

neuronxcc-2.17.194.0+d312836f/MODULE_49b8463ac655cfcc0b60+165e9558/model.hlo_module.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:192c055b31e9c41e5a0532ac236e2e0b286dbd541f739d5780d5b1d863b7c23a
+size 359350

neuronxcc-2.17.194.0+d312836f/MODULE_57b107bd0499cc4986ac+63ea7cf2/model.hlo_module.pb

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a914031f26bb159d35140c96a1bdb74370696834acf754811e40b142622a02a4
+oid sha256:e7cb531f40791a4bc8781299062f1dff17030241665ab6a7f42cb8e3161285ec
 size 68277

neuronxcc-2.17.194.0+d312836f/MODULE_57b107bd0499cc4986ac+63ea7cf2/model.neff

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c868ec53737adc2fe50d3708f1a3cedfc2e769e9e5d707555376ee2a9a153028
+oid sha256:92cc1588cd47c084f1bb989150b17e6c000400dc5fd117a11ffcc67738121dd9
 size 257024

neuronxcc-2.17.194.0+d312836f/MODULE_57b107bd0499cc4986ac+63ea7cf2/wrapped_neff.hlo

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9eee4b03603e48bfdd262029efc30ecd40c1dbfcb10b94f627cf4e6696aa82f6
+oid sha256:64b9b899e22590f54644e24b7099b1d575db60ae12f2145dee479fe538013a36
 size 268322

neuronxcc-2.17.194.0+d312836f/MODULE_64950c85776a119cdf83+b001bea4/model.hlo_module.pb

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d161deceb48f8f73a0aa6227a14f23c73b193b8c2f54ec986673930820610f16
+oid sha256:4c72a13fb767a6fe25407888d3085e20d62771714a0ae5df978f373d454fa1c6
 size 82456

neuronxcc-2.17.194.0+d312836f/MODULE_64950c85776a119cdf83+b001bea4/model.neff

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:06018e131b181177574a276f6294a56b20726b844088fb5bc1b9799f8bc4eef5
+oid sha256:7668d4bd5f98b2af1e738ec0794a4c2767b5d87a34e0b4ad8f63bd27074586ea
 size 420864

neuronxcc-2.17.194.0+d312836f/MODULE_6708f635794756c6254f+165e9558/compile_flags.json

@@ -0,0 +1 @@
+"--auto-cast=none --model-type=transformer --tensorizer-options='--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-dge-dma --vectorize-strided-dma ' -O2  --internal-num-neuroncores-per-sengine=1 --logfile=/tmp/nxd_model/token_generation_model/_tp0_bk0/log-neuron-cc.txt --enable-internal-neff-wrapper"

neuronxcc-2.17.194.0+d312836f/MODULE_6708f635794756c6254f+165e9558/model.hlo_module.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4fef8e7bcc22a6722ea6c0ab45ad76f020afc8ad127b2d6f54bbd0574cc20225
+size 359350

neuronxcc-2.17.194.0+d312836f/MODULE_6708f635794756c6254f+165e9558/model.neff

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7aa5446a880f3b507535c535e296e9aeca2e98a7c65bf0d3c920badaefd06346
+size 779264

neuronxcc-2.17.194.0+d312836f/MODULE_6708f635794756c6254f+165e9558/wrapped_neff.hlo

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:789ab784ae3447d5c7ce4adbd584ec9100168fac33c2c005070a0b28ee6e7a17
+size 799726

neuronxcc-2.17.194.0+d312836f/MODULE_6bb5680c622866b3b45b+63ea7cf2/model.hlo_module.pb

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:12aceb27cdf723d91c999dfa8673d18f8cdabe145b276c11445d19e86a5f5145
+oid sha256:731053ba55bcb60ea2044a7fa7431cc686e531e8750bd9abb746b7f014a2aee7
 size 79431