Synchronizing local compiler cache.

.gitattributes

@@ -16954,3 +16954,6 @@ neuronxcc-2.21.33363.0+82129205/MODULE_eae46774cbe298377c6c+afeb10ce/model.neff
 neuronxcc-2.21.33363.0+82129205/MODULE_eae46774cbe298377c6c+afeb10ce/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
 neuronxcc-2.21.33363.0+82129205/MODULE_af20785e9b34e9d03c3e+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text
 neuronxcc-2.21.33363.0+82129205/MODULE_af20785e9b34e9d03c3e+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.21.33363.0+82129205/MODULE_2aaf243526ada61e5bee+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.21.33363.0+82129205/MODULE_2aaf243526ada61e5bee+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
+neuronxcc-2.21.33363.0+82129205/MODULE_b7d3cb9e891fb8dc1a09+24129607/model.neff filter=lfs diff=lfs merge=lfs -text

neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/93babc5e7910f92d13b94017903ee1bc04af26270dfc2c28a995f53a81452a8c/31be4f6882236369c905.json

@@ -0,0 +1,58 @@
+{
+  "_entry_class": "SingleModelCacheEntry",
+  "_model_id": "ibm-granite/granite-3.3-8b-instruct",
+  "_task": "text-generation",
+  "architectures": [
+    "GraniteForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "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": 1,
+    "capacity_factor": null,
+    "checkpoint_id": "ibm-granite/granite-3.3-8b-instruct",
+    "checkpoint_revision": "51dd4bc2ade4059a6bd87649d68aa11e4fb2529b",
+    "continuous_batching": false,
+    "ep_degree": 1,
+    "fused_qkv": true,
+    "glu_mlp": true,
+    "local_ranks_size": 8,
+    "max_batch_size": 1,
+    "max_context_length": 16384,
+    "max_topk": 256,
+    "n_active_tokens": 16384,
+    "neuronxcc_version": "2.21.33363.0+82129205",
+    "on_device_sampling": true,
+    "optimum_neuron_version": "0.4.6.dev2",
+    "output_logits": false,
+    "pp_degree": 1,
+    "sequence_length": 16384,
+    "speculation_length": 0,
+    "start_rank_id": 0,
+    "target": "trn1",
+    "torch_dtype": "bfloat16",
+    "tp_degree": 8
+  },
+  "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": 49159
+}

neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev2/granite/ibm-granite/granite-3.3-8b-instruct/31be4f6882236369c905.json

@@ -0,0 +1,58 @@
+{
+  "_entry_class": "SingleModelCacheEntry",
+  "_model_id": "ibm-granite/granite-3.3-8b-instruct",
+  "_task": "text-generation",
+  "architectures": [
+    "GraniteForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "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": 1,
+    "capacity_factor": null,
+    "checkpoint_id": "ibm-granite/granite-3.3-8b-instruct",
+    "checkpoint_revision": "51dd4bc2ade4059a6bd87649d68aa11e4fb2529b",
+    "continuous_batching": false,
+    "ep_degree": 1,
+    "fused_qkv": true,
+    "glu_mlp": true,
+    "local_ranks_size": 8,
+    "max_batch_size": 1,
+    "max_context_length": 16384,
+    "max_topk": 256,
+    "n_active_tokens": 16384,
+    "neuronxcc_version": "2.21.33363.0+82129205",
+    "on_device_sampling": true,
+    "optimum_neuron_version": "0.4.6.dev2",
+    "output_logits": false,
+    "pp_degree": 1,
+    "sequence_length": 16384,
+    "speculation_length": 0,
+    "start_rank_id": 0,
+    "target": "trn1",
+    "torch_dtype": "bfloat16",
+    "tp_degree": 8
+  },
+  "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": 49159
+}

neuronxcc-2.21.33363.0+82129205/MODULE_2aaf243526ada61e5bee+a02c3a36/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/_tp0_bk0/log-neuron-cc.txt", "--enable-internal-neff-wrapper"]

neuronxcc-2.21.33363.0+82129205/MODULE_2aaf243526ada61e5bee+a02c3a36/model.hlo_module.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c5b85c49fc2a43c30117b0d0079b85b009fec8911d893cc33ea8336021603373
+size 1099351

neuronxcc-2.21.33363.0+82129205/MODULE_2aaf243526ada61e5bee+a02c3a36/model.neff

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9caa6a2fa93fcc854483e7515e78a537b6997b669add138a814389261f38e8a1
+size 4568064

neuronxcc-2.21.33363.0+82129205/MODULE_2aaf243526ada61e5bee+a02c3a36/wrapped_neff.hlo

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:48376737184665cbd1de6c90126f031a68fba70a8454e79f7d097a487cc13538
+size 4752683

neuronxcc-2.21.33363.0+82129205/MODULE_b7d3cb9e891fb8dc1a09+24129607/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/_tp0_bk0/log-neuron-cc.txt"]

neuronxcc-2.21.33363.0+82129205/MODULE_b7d3cb9e891fb8dc1a09+24129607/model.hlo_module.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:96451f16974db5ea30754b9c95cd5b407a881c3d349c624121d89bcfa8ed67dc
+size 922776

neuronxcc-2.21.33363.0+82129205/MODULE_b7d3cb9e891fb8dc1a09+24129607/model.neff

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5742b1459bdc582b13e52e20bee5974db09587c6bfaf8c82549240f656cf6819
+size 16671744