Synchronizing local compiler cache.

.gitattributes

@@ -9485,3 +9485,11 @@ neuronxcc-2.19.8089.0+8ab9f450/MODULE_d6446847f9a3ac00e4a9+a9d440f5/wrapped_neff
 neuronxcc-2.19.8089.0+8ab9f450/MODULE_32a2ae3a92f25d90edd5+a9d440f5/model.neff filter=lfs diff=lfs merge=lfs -text
 neuronxcc-2.19.8089.0+8ab9f450/MODULE_32a2ae3a92f25d90edd5+a9d440f5/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
 neuronxcc-2.19.8089.0+8ab9f450/MODULE_f98e3a46937c988033fa+ed72d204/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.19.8089.0+8ab9f450/MODULE_0ca0c4ac0634e856a577+431f5505/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.19.8089.0+8ab9f450/MODULE_0e04345b3b01128e3e3e+a9d440f5/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.19.8089.0+8ab9f450/MODULE_0e04345b3b01128e3e3e+a9d440f5/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.19.8089.0+8ab9f450/MODULE_3708e344e3b6526d9b66+ed72d204/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.19.8089.0+8ab9f450/MODULE_a3f54bd0729e7d8f986e+a9d440f5/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.19.8089.0+8ab9f450/MODULE_a3f54bd0729e7d8f986e+a9d440f5/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.19.8089.0+8ab9f450/MODULE_d999581831e3c2c8d5f5+431f5505/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.19.8089.0+8ab9f450/MODULE_e10842bf0119ff0ac35c+ed72d204/model.neff filter=lfs diff=lfs merge=lfs -text

neuronxcc-2.19.8089.0+8ab9f450/0_REGISTRY/0.3.0.dev5/granite/ibm-granite/granite-3.1-2b-instruct/f38b0025b4fc5af10e3d.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": 1,
+    "capacity_factor": null,
+    "cc_pipeline_tiling_factor": 2,
+    "checkpoint_id": "ibm-granite/granite-3.1-2b-instruct",
+    "checkpoint_revision": "bbc2aed595bd38bd770263dc3ab831db9794441d",
+    "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": 8,
+    "logical_nc_config": 1,
+    "max_batch_size": 1,
+    "max_context_length": 4096,
+    "max_topk": 256,
+    "mlp_kernel_enabled": false,
+    "mlp_kernel_fuse_residual_add": false,
+    "n_active_tokens": 4096,
+    "neuronxcc_version": "2.19.8089.0+8ab9f450",
+    "num_cores_per_group": 1,
+    "on_device_sampling": true,
+    "optimum_neuron_version": "0.3.0.dev5",
+    "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": 8,
+    "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-2.19.8089.0+8ab9f450/MODULE_0e04345b3b01128e3e3e+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-2.19.8089.0+8ab9f450/MODULE_0e04345b3b01128e3e3e+a9d440f5/model.hlo_module.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:69a1c76177a8bd44d67ae30d3d56ef33824c49fca4af6b8bb195593d80ad64a3
+size 1001866

neuronxcc-2.19.8089.0+8ab9f450/MODULE_0e04345b3b01128e3e3e+a9d440f5/model.neff

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:244bfb909145753b952fd30b570333b1a17bd2e096906d99e18f204f734e56ab
+size 1352704

neuronxcc-2.19.8089.0+8ab9f450/MODULE_0e04345b3b01128e3e3e+a9d440f5/wrapped_neff.hlo

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6b149affbc4852de3d028ec6588d539c74037ab03e7b88181e8f299c4c846169
+size 1526337

neuronxcc-2.19.8089.0+8ab9f450/MODULE_3708e344e3b6526d9b66+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-2.19.8089.0+8ab9f450/MODULE_3708e344e3b6526d9b66+ed72d204/model.hlo_module.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5fab2b8b80e16f50eb4d35c85c583ad7d68fe423cf7d664e84d7d1b374617938
+size 872428

neuronxcc-2.19.8089.0+8ab9f450/MODULE_3708e344e3b6526d9b66+ed72d204/model.neff

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8ad80d1eec091af164f873eda9244b3a6d15db21f0994a89533ee306f6fb1d16
+size 17767424

neuronxcc-2.19.8089.0+8ab9f450/MODULE_d999581831e3c2c8d5f5+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-2.19.8089.0+8ab9f450/MODULE_d999581831e3c2c8d5f5+431f5505/model.hlo_module.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:19fcc4af31fe54be611baf1f5e7238206a274892a932ca0d6cb43b649a5db3c3
+size 172882

neuronxcc-2.19.8089.0+8ab9f450/MODULE_d999581831e3c2c8d5f5+431f5505/model.neff

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:97310f85027084087082cc0d569e87c12ff7c23e75595696bd674dd84f4cbeb9
+size 769024