neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.3.dev1/phi3/microsoft/Phi-3-mini-4k-instruct/ff379bb81df2113b7f88.json

{
  "_entry_class": "SingleModelCacheEntry",
  "_model_id": "microsoft/Phi-3-mini-4k-instruct",
  "_task": "text-generation",
  "architectures": [
    "Phi3ForCausalLM"
  ],
  "attention_bias": false,
  "attention_dropout": 0.0,
  "auto_map": {
    "AutoConfig": "configuration_phi3.Phi3Config",
    "AutoModelForCausalLM": "modeling_phi3.Phi3ForCausalLM"
  },
  "dtype": "bfloat16",
  "embd_pdrop": 0.0,
  "hidden_act": "silu",
  "hidden_size": 3072,
  "initializer_range": 0.02,
  "intermediate_size": 8192,
  "max_position_embeddings": 4096,
  "model_type": "phi3",
  "neuron": {
    "_serialized_key": "NxDNeuronConfig",
    "batch_size": 1,
    "capacity_factor": null,
    "checkpoint_id": "microsoft/Phi-3-mini-4k-instruct",
    "checkpoint_revision": "0a67737cc96d2554230f90338b163bc6380a2a85",
    "continuous_batching": false,
    "ep_degree": 1,
    "fused_qkv": true,
    "glu_mlp": true,
    "local_ranks_size": 2,
    "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": "0.4.3.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": 32,
  "num_key_value_heads": 32,
  "original_max_position_embeddings": 4096,
  "partial_rotary_factor": 1.0,
  "resid_pdrop": 0.0,
  "rms_norm_eps": 1e-05,
  "rope_scaling": null,
  "rope_theta": 10000.0,
  "sliding_window": 2047,
  "tie_word_embeddings": false,
  "use_cache": true,
  "vocab_size": 32064
}