Synchronizing local compiler cache.

.gitattributes

@@ -17888,3 +17888,6 @@ neuronxcc-2.26.6360.0+6f180f47/MODULE_f5812595797c01ed0ee2+12dd9119/model.neff f
 neuronxcc-2.26.6360.0+6f180f47/MODULE_12bd914b7a7537225235+84337dd9/model.neff filter=lfs diff=lfs merge=lfs -text
 neuronxcc-2.26.6360.0+6f180f47/MODULE_12bd914b7a7537225235+84337dd9/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
 neuronxcc-2.26.6360.0+6f180f47/MODULE_86789c48ed32c06eeb11+12dd9119/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.26.6360.0+6f180f47/MODULE_3ce04bd396c2fbfd26e7+12dd9119/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.26.6360.0+6f180f47/MODULE_db597faa696986bab6e3+84337dd9/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.26.6360.0+6f180f47/MODULE_db597faa696986bab6e3+84337dd9/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text

neuronxcc-2.26.6360.0+6f180f47/0_REGISTRY/0.4.7.dev0/6e4d675089b03121ef48dedefef342d87e80260dd41c0bf51aa06adbcc2331f6/c9f794403303d1b75f50.json

@@ -0,0 +1,59 @@
+{
+  "_entry_class": "SingleModelCacheEntry",
+  "_model_id": "ibm-granite/granite-3.1-8b-instruct",
+  "_task": "text-generation",
+  "architectures": [
+    "GraniteForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.1,
+  "attention_multiplier": 0.0078125,
+  "dtype": "bfloat16",
+  "embedding_multiplier": 12.0,
+  "hidden_act": "silu",
+  "hidden_size": 4096,
+  "initializer_range": 0.02,
+  "intermediate_size": 12800,
+  "logits_scaling": 16.0,
+  "max_position_embeddings": 131072,
+  "mlp_bias": false,
+  "model_type": "granite",
+  "neuron": {
+    "_serialized_key": "NxDNeuronConfig",
+    "batch_size": 4,
+    "capacity_factor": null,
+    "checkpoint_id": "ibm-granite/granite-3.1-8b-instruct",
+    "checkpoint_revision": "4009206d5fc95d2e65a7b7633e159d6e97e25d35",
+    "continuous_batching": true,
+    "ep_degree": 1,
+    "fused_qkv": true,
+    "glu_mlp": true,
+    "local_ranks_size": 2,
+    "max_batch_size": 4,
+    "max_context_length": 4096,
+    "max_topk": 256,
+    "n_active_tokens": 4096,
+    "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": 4096,
+    "speculation_length": 0,
+    "start_rank_id": 0,
+    "target": "trn1",
+    "torch_dtype": "bfloat16",
+    "tp_degree": 2
+  },
+  "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": 10000000.0,
+  "tie_word_embeddings": true,
+  "use_cache": true,
+  "vocab_size": 49155
+}

neuronxcc-2.26.6360.0+6f180f47/MODULE_3ce04bd396c2fbfd26e7+12dd9119/compile_flags.json

@@ -0,0 +1 @@
+["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--verbose=35", "--logfile=/tmp/nxd_model/chunked_prefill/_tp0_bk0/log-neuron-cc.txt"]

neuronxcc-2.26.6360.0+6f180f47/MODULE_3ce04bd396c2fbfd26e7+12dd9119/model.hlo_module.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:962ba8df0b9e409defd0df1b82dc5b419563aa846418b11017e614e91e155ad9
+size 6020617

neuronxcc-2.26.6360.0+6f180f47/MODULE_3ce04bd396c2fbfd26e7+12dd9119/model.neff

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e19c46c250293e06e4f5cbd7fbd63fb626f120486c2fed497f4215e6e668b5a4
+size 4639744

neuronxcc-2.26.6360.0+6f180f47/MODULE_db597faa696986bab6e3+84337dd9/compile_flags.json

@@ -0,0 +1 @@
+["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--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_db597faa696986bab6e3+84337dd9/model.hlo_module.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:91d30100fa13353c84b2b116e0a8bcb30af314faa6e8c74e55c062fbb38966b2
+size 4353732

neuronxcc-2.26.6360.0+6f180f47/MODULE_db597faa696986bab6e3+84337dd9/model.neff

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1318cc535c5e2355c7be8be01cc02e07b0eb28c1332f37103b958faf4cdabd93
+size 12801024

neuronxcc-2.26.6360.0+6f180f47/MODULE_db597faa696986bab6e3+84337dd9/wrapped_neff.hlo

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9535bbc8d9c848418aec65b4a69049e019c04fbd55c90fbdc48baaf88f917b36
+size 12985483