quantization_config.json

{
  "bits": 4,
  "data_type": "int",
  "group_size": 32,
  "sym": true,
  "batch_size": 1,
  "iters": 1000,
  "nsamples": 512,
  "seqlen": 4096,
  "autoround_version": "0.14.2",
  "block_name_to_quantize": "model.language_model.layers",
  "quant_method": "auto-round",
  "packing_format": "auto_round:auto_gptq",
  "extra_config": {
    ".*mtp.*": {
      "data_type": "bfloat16"
    },
    ".*mtp\\.fc.*": {
      "data_type": "bfloat16"
    }
  }
}