Synchronizing local compiler cache.

.gitattributes

@@ -11209,3 +11209,6 @@ neuronxcc-2.21.18209.0+043b1bf7/MODULE_ee5a280bc35ab3275432+1fca3866/model.neff
 neuronxcc-2.21.18209.0+043b1bf7/MODULE_ee5a280bc35ab3275432+1fca3866/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
 neuronxcc-2.21.18209.0+043b1bf7/MODULE_f189f67fe2f9674d0690+ed72d204/model.neff filter=lfs diff=lfs merge=lfs -text
 neuronxcc-2.21.18209.0+043b1bf7/MODULE_f359722ac2fcd615a216+c2248236/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.21.18209.0+043b1bf7/MODULE_0af86441e7bd16c9ea64+a9d440f5/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.21.18209.0+043b1bf7/MODULE_0af86441e7bd16c9ea64+a9d440f5/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.21.18209.0+043b1bf7/MODULE_ab9f1751fa5befc35fbb+ed72d204/model.neff filter=lfs diff=lfs merge=lfs -text

neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.3.1.dev7/granite/ibm-granite/granite-3.1-2b-instruct/f85d94123356883bc6c3.json

@@ -0,0 +1,59 @@
+{
+  "_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",
+    "batch_size": 1,
+    "capacity_factor": null,
+    "checkpoint_id": "ibm-granite/granite-3.1-2b-instruct",
+    "checkpoint_revision": "bbc2aed595bd38bd770263dc3ab831db9794441d",
+    "continuous_batching": false,
+    "enable_bucketing": false,
+    "ep_degree": 1,
+    "fused_qkv": true,
+    "glu_mlp": true,
+    "local_ranks_size": 2,
+    "logical_nc_config": 1,
+    "max_batch_size": 1,
+    "max_context_length": 4096,
+    "max_topk": 256,
+    "n_active_tokens": 4096,
+    "neuronxcc_version": "2.21.18209.0+043b1bf7",
+    "on_device_sampling": true,
+    "optimum_neuron_version": "0.3.1.dev7",
+    "output_logits": false,
+    "pp_degree": 1,
+    "sequence_length": 4096,
+    "speculation_length": 0,
+    "start_rank_id": 0,
+    "target": null,
+    "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": 5000000.0,
+  "tie_word_embeddings": true,
+  "use_cache": true,
+  "vocab_size": 49155
+}

neuronxcc-2.21.18209.0+043b1bf7/MODULE_0af86441e7bd16c9ea64+a9d440f5/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", "--logfile=/tmp/nxd_model/token_generation_model/_tp0_bk0/log-neuron-cc.txt", "--enable-internal-neff-wrapper"]

neuronxcc-2.21.18209.0+043b1bf7/MODULE_0af86441e7bd16c9ea64+a9d440f5/model.hlo_module.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8e586fdee55272731cb97293d5a59f0c2b1e83962ffd382ee5de351fb9a10a26
+size 1071461

neuronxcc-2.21.18209.0+043b1bf7/MODULE_0af86441e7bd16c9ea64+a9d440f5/model.neff

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ca4198f9371f2cab5311514dc434e8b5c7ce7182712d8247b6e94231d872346b
+size 4434944

neuronxcc-2.21.18209.0+043b1bf7/MODULE_0af86441e7bd16c9ea64+a9d440f5/wrapped_neff.hlo

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4e7332c55ba62d21019e2449099ff073173f7590767912ef48091f73adab86a8
+size 4619266

neuronxcc-2.21.18209.0+043b1bf7/MODULE_ab9f1751fa5befc35fbb+ed72d204/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", "--logfile=/tmp/nxd_model/context_encoding_model/_tp0_bk0/log-neuron-cc.txt"]

neuronxcc-2.21.18209.0+043b1bf7/MODULE_ab9f1751fa5befc35fbb+ed72d204/model.hlo_module.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cebec0aa04c996d78d53d9f546e3bac3389c7355c40a51c015c27d82566537e2
+size 961456

neuronxcc-2.21.18209.0+043b1bf7/MODULE_ab9f1751fa5befc35fbb+ed72d204/model.neff

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f9f126645dc25cb24e4827660ba5e21516889dd6a6420679c61db293c1eabb7e
+size 3769344