quantization_config.json

{
  "bits": 2,
  "group_size": 64,
  "sym": true,
  "data_type": "int",
  "enable_quanted_input": true,
  "enable_minmax_tuning": true,
  "seqlen": 2048,
  "batch_size": 8,
  "scale_dtype": "torch.float16",
  "lr": 0.0005,
  "minmax_lr": 0.0005,
  "gradient_accumulate_steps": 1,
  "iters": 2000,
  "amp": true,
  "nsamples": 1024,
  "low_gpu_mem_usage": true,
  "to_quant_block_names": "model.layers",
  "enable_norm_bias_tuning": false,
  "dataset": "NeelNanda/pile-10k",
  "autoround_version": "0.4.2",
  "quant_method": "intel/auto-round",
  "backend": "auto_round:gptq:exllamav2"
}