{
  "model_type": "llava",
  "quantization": "q4f16_1",
  "model_config": {
    "image_token_index": 32011,
    "text_config": {
      "model_type": "phi3",
      "hidden_size": 3072,
      "vocab_size": 32012,
      "num_hidden_layers": 32,
      "num_attention_heads": 32,
      "intermediate_size": 8192,
      "rms_norm_eps": 1e-05,
      "num_key_value_heads": 32,
      "position_embedding_base": 10000.0,
      "context_window_size": 4096,
      "prefill_chunk_size": 2048,
      "head_dim": 96,
      "tensor_parallel_shards": 1,
      "max_batch_size": 1,
      "kwargs": {
        "_name_or_path": "Phi-3-mini-4k-instruct",
        "architectures": [
          "Phi3ForCausalLM"
        ],
        "embd_pdrop": 0.0,
        "eos_token_id": 32000,
        "original_max_position_embeddings": 4096,
        "pad_token_id": 32000,
        "resid_pdrop": 0.0,
        "sliding_window": 2047,
        "torch_dtype": "bfloat16",
        "attention_dropout": 0.0,
        "auto_map": {
          "AutoConfig": "configuration_phi3.Phi3Config",
          "AutoModelForCausalLM": "modeling_phi3.Phi3ForCausalLM"
        },
        "bos_token_id": 1,
        "hidden_act": "silu",
        "initializer_range": 0.02,
        "rope_scaling": null,
        "tie_word_embeddings": false,
        "transformers_version": "4.39.3",
        "use_cache": true
      }
    },
    "vision_config": {
      "hidden_size": 768,
      "image_size": 224,
      "intermediate_size": 3072,
      "num_attention_heads": 8,
      "num_hidden_layers": 12,
      "patch_size": 16,
      "projection_dim": 512,
      "vocab_size": 32012,
      "num_channels": 3,
      "layer_norm_eps": 1e-06,
      "kwargs": {
        "model_type": "clip_vision_model"
      }
    },
    "vocab_size": 32012,
    "context_window_size": 768,
    "sliding_window_size": -1,
    "prefill_chunk_size": 2048,
    "tensor_parallel_shards": 1,
    "max_batch_size": 80,
    "text_architecture": "Phi3ForCausalLM"
  },
  "vocab_size": 32012,
  "context_window_size": 768,
  "sliding_window_size": -1,
  "prefill_chunk_size": 2048,
  "attention_sink_size": -1,
  "tensor_parallel_shards": 1,
  "mean_gen_len": 128,
  "max_gen_len": 512,
  "shift_fill_factor": 0.3,
  "temperature": 0,
  "presence_penalty": 0.0,
  "frequency_penalty": 0.0,
  "repetition_penalty": 1.0,
  "top_p": 0.95,
  "conv_template": {
    "name": "phi-3",
    "system_template": "<|system|>\n{system_message}",
    "system_message": "You are a helpful digital assistant. Please provide safe, ethical and accurate information to the user.",
    "system_prefix_token_ids": [
      1
    ],
    "add_role_after_system_message": true,
    "roles": {
      "user": "<|user|>",
      "assistant": "<|assistant|>"
    },
    "role_templates": {
      "user": "{user_message}",
      "assistant": "{assistant_message}",
      "tool": "{tool_message}"
    },
    "messages": [],
    "seps": [
      "<|end|>\n"
    ],
    "role_content_sep": "\n",
    "role_empty_sep": "\n",
    "stop_str": [
      "<|endoftext|>"
    ],
    "stop_token_ids": [
      32000,
      32001,
      32007
    ],
    "function_string": "",
    "use_function_calling": false
  },
  "pad_token_id": 0,
  "bos_token_id": 1,
  "eos_token_id": 32000,
  "tokenizer_files": [
    "tokenizer.model",
    "tokenizer.json",
    "added_tokens.json",
    "tokenizer_config.json"
  ],
  "version": "0.1.0"
}