quantization_config.json

{
  "bits": 4,
  "group_size": 128,
  "sym": true,
  "data_type": "int",
  "seqlen": 512,
  "batch_size": 1,
  "scale_dtype": "None",
  "gradient_accumulate_steps": 8,
  "iters": 100,
  "to_quant_block_names": "model.language_model.layers",
  "dataset": "/data/fengdahu/tmp/dataset_calib1208/mobimind_calibration_20k.json",
  "autoround_version": "0.9.2",
  "provider": "auto-round",
  "quant_method": "awq",
  "zero_point": false,
  "version": "gemm",
  "modules_to_not_convert": [
    "blocks.0.attn.qkv",
    "blocks.0.attn.proj",
    "blocks.0.mlp.linear_fc1",
    "blocks.0.mlp.linear_fc2",
    "blocks.1.attn.qkv",
    "blocks.1.attn.proj",
    "blocks.1.mlp.linear_fc1",
    "blocks.1.mlp.linear_fc2",
    "blocks.2.attn.qkv",
    "blocks.2.attn.proj",
    "blocks.2.mlp.linear_fc1",
    "blocks.2.mlp.linear_fc2",
    "blocks.3.attn.qkv",
    "blocks.3.attn.proj",
    "blocks.3.mlp.linear_fc1",
    "blocks.3.mlp.linear_fc2",
    "blocks.4.attn.qkv",
    "blocks.4.attn.proj",
    "blocks.4.mlp.linear_fc1",
    "blocks.4.mlp.linear_fc2",
    "blocks.5.attn.qkv",
    "blocks.5.attn.proj",
    "blocks.5.mlp.linear_fc1",
    "blocks.5.mlp.linear_fc2",
    "blocks.6.attn.qkv",
    "blocks.6.attn.proj",
    "blocks.6.mlp.linear_fc1",
    "blocks.6.mlp.linear_fc2",
    "blocks.7.attn.qkv",
    "blocks.7.attn.proj",
    "blocks.7.mlp.linear_fc1",
    "blocks.7.mlp.linear_fc2",
    "blocks.8.attn.qkv",
    "blocks.8.attn.proj",
    "blocks.8.mlp.linear_fc1",
    "blocks.8.mlp.linear_fc2",
    "blocks.9.attn.qkv",
    "blocks.9.attn.proj",
    "blocks.9.mlp.linear_fc1",
    "blocks.9.mlp.linear_fc2",
    "blocks.10.attn.qkv",
    "blocks.10.attn.proj",
    "blocks.10.mlp.linear_fc1",
    "blocks.10.mlp.linear_fc2",
    "blocks.11.attn.qkv",
    "blocks.11.attn.proj",
    "blocks.11.mlp.linear_fc1",
    "blocks.11.mlp.linear_fc2",
    "blocks.12.attn.qkv",
    "blocks.12.attn.proj",
    "blocks.12.mlp.linear_fc1",
    "blocks.12.mlp.linear_fc2",
    "blocks.13.attn.qkv",
    "blocks.13.attn.proj",
    "blocks.13.mlp.linear_fc1",
    "blocks.13.mlp.linear_fc2",
    "blocks.14.attn.qkv",
    "blocks.14.attn.proj",
    "blocks.14.mlp.linear_fc1",
    "blocks.14.mlp.linear_fc2",
    "blocks.15.attn.qkv",
    "blocks.15.attn.proj",
    "blocks.15.mlp.linear_fc1",
    "blocks.15.mlp.linear_fc2",
    "blocks.16.attn.qkv",
    "blocks.16.attn.proj",
    "blocks.16.mlp.linear_fc1",
    "blocks.16.mlp.linear_fc2",
    "blocks.17.attn.qkv",
    "blocks.17.attn.proj",
    "blocks.17.mlp.linear_fc1",
    "blocks.17.mlp.linear_fc2",
    "blocks.18.attn.qkv",
    "blocks.18.attn.proj",
    "blocks.18.mlp.linear_fc1",
    "blocks.18.mlp.linear_fc2",
    "blocks.19.attn.qkv",
    "blocks.19.attn.proj",
    "blocks.19.mlp.linear_fc1",
    "blocks.19.mlp.linear_fc2",
    "blocks.20.attn.qkv",
    "blocks.20.attn.proj",
    "blocks.20.mlp.linear_fc1",
    "blocks.20.mlp.linear_fc2",
    "blocks.21.attn.qkv",
    "blocks.21.attn.proj",
    "blocks.21.mlp.linear_fc1",
    "blocks.21.mlp.linear_fc2",
    "blocks.22.attn.qkv",
    "blocks.22.attn.proj",
    "blocks.22.mlp.linear_fc1",
    "blocks.22.mlp.linear_fc2",
    "blocks.23.attn.qkv",
    "blocks.23.attn.proj",
    "blocks.23.mlp.linear_fc1",
    "blocks.23.mlp.linear_fc2",
    "merger.linear_fc1",
    "merger.linear_fc2",
    "deepstack_merger_list.0.linear_fc1",
    "deepstack_merger_list.0.linear_fc2",
    "deepstack_merger_list.1.linear_fc1",
    "deepstack_merger_list.1.linear_fc2",
    "deepstack_merger_list.2.linear_fc1",
    "deepstack_merger_list.2.linear_fc2",
    "lm_head",
    "model.visual.deepstack_merger_list",
    "model.visual.blocks"
  ]
}