Synchronizing local compiler cache.

.gitattributes

@@ -13613,3 +13613,6 @@ neuronxcc-2.21.33363.0+82129205/MODULE_a36a9eb929c263812fff+6e4949b4/model.neff
 neuronxcc-2.21.33363.0+82129205/MODULE_9c081fe8e4d6036465c0+283df001/model.neff filter=lfs diff=lfs merge=lfs -text
 neuronxcc-2.21.33363.0+82129205/MODULE_e6ccdb863943230b38f7+2dde74c7/model.neff filter=lfs diff=lfs merge=lfs -text
 neuronxcc-2.21.33363.0+82129205/MODULE_e6ccdb863943230b38f7+2dde74c7/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.21.33363.0+82129205/MODULE_2855e6e20fc579be1a4e+ac10809c/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.21.33363.0+82129205/MODULE_2855e6e20fc579be1a4e+ac10809c/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.21.33363.0+82129205/MODULE_fd36be5d3561ffaa3ebf+a32116a7/model.neff filter=lfs diff=lfs merge=lfs -text

neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.3.dev2/llama/meta-llama/Llama-3.2-1B/1afc1cf8d7f8f7b0a6a6.json

@@ -0,0 +1,62 @@
+{
+  "_entry_class": "SingleModelCacheEntry",
+  "_model_id": "meta-llama/Llama-3.2-1B",
+  "_task": "text-generation",
+  "architectures": [
+    "LlamaForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "dtype": "bfloat16",
+  "head_dim": 64,
+  "hidden_act": "silu",
+  "hidden_size": 2048,
+  "initializer_range": 0.02,
+  "intermediate_size": 8192,
+  "max_position_embeddings": 131072,
+  "mlp_bias": false,
+  "model_type": "llama",
+  "neuron": {
+    "_serialized_key": "NxDNeuronConfig",
+    "batch_size": 1,
+    "capacity_factor": null,
+    "checkpoint_id": "meta-llama/Llama-3.2-1B",
+    "checkpoint_revision": "4e20de362430cd3b72f300e6b0f18e50e7166e08",
+    "continuous_batching": false,
+    "ep_degree": 1,
+    "fused_qkv": true,
+    "glu_mlp": true,
+    "local_ranks_size": 4,
+    "max_batch_size": 1,
+    "max_context_length": 4096,
+    "max_topk": 256,
+    "n_active_tokens": 4096,
+    "neuronxcc_version": "2.21.33363.0+82129205",
+    "on_device_sampling": true,
+    "optimum_neuron_version": "0.4.3.dev2",
+    "output_logits": false,
+    "pp_degree": 1,
+    "sequence_length": 4096,
+    "speculation_length": 0,
+    "start_rank_id": 0,
+    "target": "trn2",
+    "torch_dtype": "bfloat16",
+    "tp_degree": 4
+  },
+  "num_attention_heads": 32,
+  "num_hidden_layers": 16,
+  "num_key_value_heads": 8,
+  "pretraining_tp": 1,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": {
+    "factor": 32.0,
+    "high_freq_factor": 4.0,
+    "low_freq_factor": 1.0,
+    "original_max_position_embeddings": 8192,
+    "rope_type": "llama3"
+  },
+  "rope_theta": 500000.0,
+  "tie_word_embeddings": true,
+  "use_cache": true,
+  "vocab_size": 128256
+}

neuronxcc-2.21.33363.0+82129205/MODULE_2855e6e20fc579be1a4e+ac10809c/compile_flags.json

@@ -0,0 +1 @@
+["--target=trn2", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=2", "--logfile=/tmp/nxd_model/token_generation/_tp0_bk0/log-neuron-cc.txt", "--enable-internal-neff-wrapper"]

neuronxcc-2.21.33363.0+82129205/MODULE_2855e6e20fc579be1a4e+ac10809c/model.hlo_module.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:98e000609db7073f994d912f1fb42f02f2670234e130f5d25bfbe327ddee0f7c
+size 1519995

neuronxcc-2.21.33363.0+82129205/MODULE_2855e6e20fc579be1a4e+ac10809c/model.neff

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9f5542f1ed54cb9dab39fc2ea1a4400b5b052caefb1fa535e9e44db60d8c035f
+size 1803264

neuronxcc-2.21.33363.0+82129205/MODULE_2855e6e20fc579be1a4e+ac10809c/wrapped_neff.hlo

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:565140f6546c221bc5ba9962b433c00809afdd71ced7ff84ad9cced3b5345632
+size 1878847

neuronxcc-2.21.33363.0+82129205/MODULE_fd36be5d3561ffaa3ebf+a32116a7/compile_flags.json

@@ -0,0 +1 @@
+["--target=trn2", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=2", "--logfile=/tmp/nxd_model/context_encoding/_tp0_bk0/log-neuron-cc.txt"]

neuronxcc-2.21.33363.0+82129205/MODULE_fd36be5d3561ffaa3ebf+a32116a7/model.hlo_module.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bf2b388fdc02ddbccaf0d0bbd1be4c84cd48fea4e98e7a11fb3f0bc1f55fdafd
+size 1638255

neuronxcc-2.21.33363.0+82129205/MODULE_fd36be5d3561ffaa3ebf+a32116a7/model.neff

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3f1933be025b9057be1381723b5dd4d3ced95d2b6646d409b3bc36b49d02374e
+size 18863104