mlc-chat-config.json

{
  "model_type": "asvd_qwen2",
  "quantization": "q0f16",
  "model_config": {
    "hidden_size": 3584,
    "intermediate_size": 18944,
    "num_attention_heads": 28,
    "num_hidden_layers": 28,
    "rms_norm_eps": 1e-06,
    "vocab_size": 90990,
    "position_embedding_base": 1000000.0,
    "context_window_size": 768,
    "prefill_chunk_size": 768,
    "num_key_value_heads": 4,
    "head_dim": 128,
    "tensor_parallel_shards": 1,
    "truncation_ranks": {
      "model.layers.0.mlp.gate_proj": [
        1085,
        120
      ],
      "model.layers.0.mlp.up_proj": [
        678,
        75
      ],
      "model.layers.0.self_attn.k_proj": [
        101,
        11
      ],
      "model.layers.0.self_attn.o_proj": [
        887,
        98
      ],
      "model.layers.0.self_attn.q_proj": [
        162,
        17
      ],
      "model.layers.0.self_attn.v_proj": [
        342,
        38
      ],
      "model.layers.1.mlp.down_proj": [
        2305,
        256
      ],
      "model.layers.1.mlp.gate_proj": [
        1492,
        165
      ],
      "model.layers.1.mlp.up_proj": [
        678,
        75
      ],
      "model.layers.1.self_attn.k_proj": [
        41,
        4
      ],
      "model.layers.1.self_attn.o_proj": [
        887,
        98
      ],
      "model.layers.1.self_attn.q_proj": [
        162,
        17
      ],
      "model.layers.1.self_attn.v_proj": [
        222,
        24
      ],
      "model.layers.10.mlp.gate_proj": [
        1899,
        210
      ],
      "model.layers.10.mlp.up_proj": [
        1492,
        165
      ],
      "model.layers.10.self_attn.k_proj": [
        101,
        11
      ],
      "model.layers.10.self_attn.o_proj": [
        645,
        71
      ],
      "model.layers.10.self_attn.q_proj": [
        162,
        17
      ],
      "model.layers.11.mlp.down_proj": [
        1492,
        165
      ],
      "model.layers.11.mlp.gate_proj": [
        1899,
        210
      ],
      "model.layers.11.self_attn.k_proj": [
        101,
        11
      ],
      "model.layers.11.self_attn.o_proj": [
        645,
        71
      ],
      "model.layers.11.self_attn.q_proj": [
        403,
        44
      ],
      "model.layers.11.self_attn.v_proj": [
        222,
        24
      ],
      "model.layers.12.self_attn.k_proj": [
        101,
        11
      ],
      "model.layers.12.self_attn.o_proj": [
        403,
        44
      ],
      "model.layers.12.self_attn.q_proj": [
        162,
        17
      ],
      "model.layers.12.self_attn.v_proj": [
        342,
        38
      ],
      "model.layers.13.mlp.down_proj": [
        1085,
        120
      ],
      "model.layers.13.mlp.gate_proj": [
        1492,
        165
      ],
      "model.layers.13.mlp.up_proj": [
        1492,
        165
      ],
      "model.layers.13.self_attn.k_proj": [
        101,
        11
      ],
      "model.layers.13.self_attn.o_proj": [
        645,
        71
      ],
      "model.layers.13.self_attn.q_proj": [
        403,
        44
      ],
      "model.layers.14.mlp.down_proj": [
        1085,
        120
      ],
      "model.layers.14.mlp.gate_proj": [
        1899,
        210
      ],
      "model.layers.14.self_attn.k_proj": [
        101,
        11
      ],
      "model.layers.14.self_attn.o_proj": [
        162,
        17
      ],
      "model.layers.14.self_attn.q_proj": [
        162,
        17
      ],
      "model.layers.14.self_attn.v_proj": [
        162,
        17
      ],
      "model.layers.15.mlp.down_proj": [
        1492,
        165
      ],
      "model.layers.15.mlp.gate_proj": [
        1492,
        165
      ],
      "model.layers.15.mlp.up_proj": [
        1899,
        210
      ],
      "model.layers.15.self_attn.k_proj": [
        162,
        17
      ],
      "model.layers.15.self_attn.o_proj": [
        645,
        71
      ],
      "model.layers.15.self_attn.q_proj": [
        403,
        44
      ],
      "model.layers.15.self_attn.v_proj": [
        342,
        38
      ],
      "model.layers.16.mlp.down_proj": [
        1492,
        165
      ],
      "model.layers.16.mlp.gate_proj": [
        1492,
        165
      ],
      "model.layers.16.self_attn.k_proj": [
        101,
        11
      ],
      "model.layers.16.self_attn.o_proj": [
        645,
        71
      ],
      "model.layers.16.self_attn.q_proj": [
        403,
        44
      ],
      "model.layers.17.mlp.down_proj": [
        1085,
        120
      ],
      "model.layers.17.mlp.gate_proj": [
        678,
        75
      ],
      "model.layers.17.mlp.up_proj": [
        1899,
        210
      ],
      "model.layers.17.self_attn.k_proj": [
        162,
        17
      ],
      "model.layers.17.self_attn.o_proj": [
        403,
        44
      ],
      "model.layers.17.self_attn.q_proj": [
        403,
        44
      ],
      "model.layers.17.self_attn.v_proj": [
        162,
        17
      ],
      "model.layers.18.mlp.down_proj": [
        1492,
        165
      ],
      "model.layers.18.mlp.gate_proj": [
        1492,
        165
      ],
      "model.layers.18.mlp.up_proj": [
        1492,
        165
      ],
      "model.layers.18.self_attn.k_proj": [
        162,
        17
      ],
      "model.layers.18.self_attn.o_proj": [
        403,
        44
      ],
      "model.layers.18.self_attn.q_proj": [
        162,
        17
      ],
      "model.layers.18.self_attn.v_proj": [
        342,
        38
      ],
      "model.layers.19.mlp.down_proj": [
        1899,
        210
      ],
      "model.layers.19.mlp.gate_proj": [
        1492,
        165
      ],
      "model.layers.19.mlp.up_proj": [
        1085,
        120
      ],
      "model.layers.19.self_attn.k_proj": [
        101,
        11
      ],
      "model.layers.19.self_attn.o_proj": [
        887,
        98
      ],
      "model.layers.19.self_attn.q_proj": [
        403,
        44
      ],
      "model.layers.2.mlp.gate_proj": [
        1899,
        210
      ],
      "model.layers.2.mlp.up_proj": [
        1085,
        120
      ],
      "model.layers.2.self_attn.k_proj": [
        101,
        11
      ],
      "model.layers.2.self_attn.o_proj": [
        645,
        71
      ],
      "model.layers.2.self_attn.q_proj": [
        162,
        17
      ],
      "model.layers.2.self_attn.v_proj": [
        222,
        24
      ],
      "model.layers.20.mlp.gate_proj": [
        1492,
        165
      ],
      "model.layers.20.mlp.up_proj": [
        1899,
        210
      ],
      "model.layers.20.self_attn.k_proj": [
        101,
        11
      ],
      "model.layers.20.self_attn.o_proj": [
        403,
        44
      ],
      "model.layers.20.self_attn.q_proj": [
        403,
        44
      ],
      "model.layers.20.self_attn.v_proj": [
        162,
        17
      ],
      "model.layers.21.mlp.gate_proj": [
        1899,
        210
      ],
      "model.layers.21.mlp.up_proj": [
        1899,
        210
      ],
      "model.layers.21.self_attn.k_proj": [
        101,
        11
      ],
      "model.layers.21.self_attn.o_proj": [
        645,
        71
      ],
      "model.layers.21.self_attn.q_proj": [
        403,
        44
      ],
      "model.layers.22.mlp.gate_proj": [
        2305,
        256
      ],
      "model.layers.22.mlp.up_proj": [
        1492,
        165
      ],
      "model.layers.22.self_attn.k_proj": [
        101,
        11
      ],
      "model.layers.22.self_attn.o_proj": [
        403,
        44
      ],
      "model.layers.22.self_attn.q_proj": [
        162,
        17
      ],
      "model.layers.22.self_attn.v_proj": [
        282,
        31
      ],
      "model.layers.23.mlp.gate_proj": [
        2305,
        256
      ],
      "model.layers.23.mlp.up_proj": [
        1899,
        210
      ],
      "model.layers.23.self_attn.k_proj": [
        101,
        11
      ],
      "model.layers.23.self_attn.o_proj": [
        645,
        71
      ],
      "model.layers.23.self_attn.q_proj": [
        403,
        44
      ],
      "model.layers.24.mlp.gate_proj": [
        2305,
        256
      ],
      "model.layers.24.mlp.up_proj": [
        1492,
        165
      ],
      "model.layers.24.self_attn.k_proj": [
        101,
        11
      ],
      "model.layers.24.self_attn.o_proj": [
        403,
        44
      ],
      "model.layers.24.self_attn.q_proj": [
        403,
        44
      ],
      "model.layers.25.mlp.up_proj": [
        2305,
        256
      ],
      "model.layers.25.self_attn.k_proj": [
        101,
        11
      ],
      "model.layers.25.self_attn.o_proj": [
        162,
        17
      ],
      "model.layers.25.self_attn.q_proj": [
        162,
        17
      ],
      "model.layers.25.self_attn.v_proj": [
        222,
        24
      ],
      "model.layers.26.self_attn.k_proj": [
        101,
        11
      ],
      "model.layers.26.self_attn.o_proj": [
        403,
        44
      ],
      "model.layers.26.self_attn.q_proj": [
        162,
        17
      ],
      "model.layers.26.self_attn.v_proj": [
        162,
        17
      ],
      "model.layers.27.self_attn.k_proj": [
        101,
        11
      ],
      "model.layers.27.self_attn.o_proj": [
        645,
        71
      ],
      "model.layers.27.self_attn.q_proj": [
        162,
        17
      ],
      "model.layers.27.self_attn.v_proj": [
        162,
        17
      ],
      "model.layers.3.mlp.gate_proj": [
        1899,
        210
      ],
      "model.layers.3.mlp.up_proj": [
        1492,
        165
      ],
      "model.layers.3.self_attn.k_proj": [
        101,
        11
      ],
      "model.layers.3.self_attn.o_proj": [
        887,
        98
      ],
      "model.layers.3.self_attn.q_proj": [
        162,
        17
      ],
      "model.layers.3.self_attn.v_proj": [
        282,
        31
      ],
      "model.layers.4.mlp.up_proj": [
        1899,
        210
      ],
      "model.layers.4.self_attn.k_proj": [
        101,
        11
      ],
      "model.layers.4.self_attn.o_proj": [
        403,
        44
      ],
      "model.layers.4.self_attn.q_proj": [
        162,
        17
      ],
      "model.layers.4.self_attn.v_proj": [
        162,
        17
      ],
      "model.layers.5.mlp.down_proj": [
        1899,
        210
      ],
      "model.layers.5.self_attn.k_proj": [
        162,
        17
      ],
      "model.layers.5.self_attn.o_proj": [
        162,
        17
      ],
      "model.layers.5.self_attn.q_proj": [
        162,
        17
      ],
      "model.layers.5.self_attn.v_proj": [
        101,
        11
      ],
      "model.layers.6.self_attn.k_proj": [
        41,
        4
      ],
      "model.layers.6.self_attn.o_proj": [
        645,
        71
      ],
      "model.layers.6.self_attn.q_proj": [
        162,
        17
      ],
      "model.layers.7.mlp.gate_proj": [
        2305,
        256
      ],
      "model.layers.7.self_attn.k_proj": [
        41,
        4
      ],
      "model.layers.7.self_attn.o_proj": [
        887,
        98
      ],
      "model.layers.7.self_attn.q_proj": [
        162,
        17
      ],
      "model.layers.8.mlp.gate_proj": [
        1899,
        210
      ],
      "model.layers.8.mlp.up_proj": [
        2305,
        256
      ],
      "model.layers.8.self_attn.k_proj": [
        162,
        17
      ],
      "model.layers.8.self_attn.o_proj": [
        645,
        71
      ],
      "model.layers.8.self_attn.q_proj": [
        403,
        44
      ],
      "model.layers.8.self_attn.v_proj": [
        282,
        31
      ],
      "model.layers.9.mlp.down_proj": [
        1899,
        210
      ],
      "model.layers.9.mlp.gate_proj": [
        2305,
        256
      ],
      "model.layers.9.self_attn.k_proj": [
        41,
        4
      ],
      "model.layers.9.self_attn.o_proj": [
        403,
        44
      ],
      "model.layers.9.self_attn.q_proj": [
        162,
        17
      ],
      "model.layers.9.self_attn.v_proj": [
        342,
        38
      ]
    },
    "max_batch_size": 1
  },
  "vocab_size": 90990,
  "context_window_size": 768,
  "sliding_window_size": -1,
  "prefill_chunk_size": 768,
  "attention_sink_size": -1,
  "tensor_parallel_shards": 1,
  "max_batch_size": 80,
  "mean_gen_len": 128,
  "max_gen_len": 512,
  "shift_fill_factor": 0.3,
  "temperature": 0.7,
  "repetition_penalty": 1.05,
  "top_p": 0.8,
  "conv_template": "LM",
  "pad_token_id": 151643,
  "bos_token_id": 151643,
  "eos_token_id": [
    151645,
    151643
  ],
  "tokenizer_files": [
    "tokenizer.json",
    "vocab.json",
    "merges.txt",
    "added_tokens.json",
    "tokenizer_config.json"
  ],
  "version": "0.1.0"
}