Synchronizing local compiler cache.

neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.3.dev1/granite/ibm-granite/granite-3.1-2b-instruct/d796a012004414fbb58a.json

@@ -0,0 +1,58 @@
+{
+  "_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,
+  "dtype": "bfloat16",
+  "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": 4,
+    "capacity_factor": null,
+    "checkpoint_id": "ibm-granite/granite-3.1-2b-instruct",
+    "checkpoint_revision": "bbc2aed595bd38bd770263dc3ab831db9794441d",
+    "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": "2.21.18209.0+043b1bf7",
+    "on_device_sampling": true,
+    "optimum_neuron_version": "0.4.3.dev1",
+    "output_logits": false,
+    "pp_degree": 1,
+    "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": 5000000.0,
+  "tie_word_embeddings": true,
+  "use_cache": true,
+  "vocab_size": 49155
+}