Synchronizing local compiler cache.

.gitattributes

@@ -18254,3 +18254,6 @@ neuronxcc-2.26.6360.0+6f180f47/MODULE_e83c28d0c5f743fb626e+4a27301b/wrapped_neff
 neuronxcc-2.26.6360.0+6f180f47/MODULE_1360e4f594826eb48799+4a27301b/model.neff filter=lfs diff=lfs merge=lfs -text
 neuronxcc-2.26.6360.0+6f180f47/MODULE_1360e4f594826eb48799+4a27301b/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
 neuronxcc-2.26.6360.0+6f180f47/MODULE_eed8bdba87555d9b4043+665810b0/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.26.6360.0+6f180f47/MODULE_a5036613bd3c1751af0b+665810b0/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.26.6360.0+6f180f47/MODULE_f4895f64e21175a2dd9e+4a27301b/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.26.6360.0+6f180f47/MODULE_f4895f64e21175a2dd9e+4a27301b/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text

neuronxcc-2.26.6360.0+6f180f47/0_REGISTRY/0.4.7.dev0/5b1dc9e6446dbcb9bab79e92fa65dc122cd2ae7d8bd595492762496ea51f4d64/4890e2c8179681b9ca24.json

@@ -0,0 +1,63 @@
+{
+  "_entry_class": "SingleModelCacheEntry",
+  "_model_id": "meta-llama/Llama-3.2-3B",
+  "_task": "text-generation",
+  "architectures": [
+    "LlamaForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "dtype": "bfloat16",
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 3072,
+  "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-3.2-3B",
+    "checkpoint_revision": "13afe5124825b4f3751f836b40dafda64c1ed062",
+    "continuous_batching": false,
+    "ep_degree": 1,
+    "fused_qkv": true,
+    "glu_mlp": true,
+    "local_ranks_size": 4,
+    "max_batch_size": 1,
+    "max_context_length": 16384,
+    "max_topk": 256,
+    "n_active_tokens": 16384,
+    "neuronxcc_version": "2.26.6360.0+6f180f47",
+    "on_device_sampling": true,
+    "optimum_neuron_version": "0.4.7.dev0",
+    "output_logits": false,
+    "pp_degree": 1,
+    "prefill_chunk_size": 1024,
+    "sequence_length": 16384,
+    "speculation_length": 0,
+    "start_rank_id": 0,
+    "target": "trn2",
+    "torch_dtype": "bfloat16",
+    "tp_degree": 4
+  },
+  "num_attention_heads": 24,
+  "num_hidden_layers": 28,
+  "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-2.26.6360.0+6f180f47/MODULE_a5036613bd3c1751af0b+665810b0/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", "--verbose=35", "--logfile=/tmp/nxd_model/chunked_prefill/_tp0_bk0/log-neuron-cc.txt"]

neuronxcc-2.26.6360.0+6f180f47/MODULE_a5036613bd3c1751af0b+665810b0/model.hlo_module.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cc871431cab33dfc4f8a2c77f98efa273f091d7167eaed97792eb735c154ef0e
+size 3824528

neuronxcc-2.26.6360.0+6f180f47/MODULE_a5036613bd3c1751af0b+665810b0/model.neff

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ac16cfba407a8fd5aca0a6e180eac2896905bb95f26f95e18494b866bb7db64c
+size 7927064

neuronxcc-2.26.6360.0+6f180f47/MODULE_f4895f64e21175a2dd9e+4a27301b/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", "--verbose=35", "--logfile=/tmp/nxd_model/token_generation/_tp0_bk0/log-neuron-cc.txt", "--enable-internal-neff-wrapper"]

neuronxcc-2.26.6360.0+6f180f47/MODULE_f4895f64e21175a2dd9e+4a27301b/model.hlo_module.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b11fa07528721044dfa4c5ca42d545e682901001a47aff56ed5f520152792ef6
+size 2940188

neuronxcc-2.26.6360.0+6f180f47/MODULE_f4895f64e21175a2dd9e+4a27301b/model.neff

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:58adb284416a05462c8b66b3733417ede1228217dddf3f639c696308bbb835a8
+size 3684335

neuronxcc-2.26.6360.0+6f180f47/MODULE_f4895f64e21175a2dd9e+4a27301b/wrapped_neff.hlo

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:94ab7e71b6e7f23d2e0161752e1a01e96835486eab19ec2b17f56980cc262ca5
+size 3816001