neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.2.2/phi3/microsoft/Phi-3-mini-4k-instruct/4d3e15aa937de5d88ebc.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": "microsoft/Phi-3-mini-4k-instruct--configuration_phi3.Phi3Config",
    "AutoModelForCausalLM": "microsoft/Phi-3-mini-4k-instruct--modeling_phi3.Phi3ForCausalLM"
  },
  "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",
    "async_mode": false,
    "attn_kernel_enabled": false,
    "batch_size": 4,
    "capacity_factor": null,
    "cc_pipeline_tiling_factor": 2,
    "checkpoint_id": "microsoft/Phi-3-mini-4k-instruct",
    "checkpoint_revision": "0a67737cc96d2554230f90338b163bc6380a2a85",
    "continuous_batching": true,
    "enable_bucketing": false,
    "ep_degree": 1,
    "flash_decoding_enabled": false,
    "fused_qkv": true,
    "glu_mlp": true,
    "is_chunked_prefill": false,
    "local_ranks_size": 2,
    "logical_nc_config": 1,
    "max_batch_size": 4,
    "max_context_length": 4096,
    "max_topk": 256,
    "mlp_kernel_enabled": false,
    "mlp_kernel_fuse_residual_add": false,
    "n_active_tokens": 4096,
    "neuronxcc_version": "2.17.194.0+d312836f",
    "num_cores_per_group": 1,
    "on_device_sampling": false,
    "optimum_neuron_version": "0.2.2",
    "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": 2,
    "vocab_parallel": false
  },
  "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
}