quantization_config.json

{
  "bits": 4,
  "act_bits": 32,
  "group_size": 128,
  "sym": true,
  "data_type": "int",
  "enable_quanted_input": true,
  "enable_minmax_tuning": true,
  "seqlen": 2048,
  "batch_size": 8,
  "scale_dtype": "torch.float16",
  "lr": 0.005,
  "minmax_lr": 0.005,
  "gradient_accumulate_steps": 1,
  "iters": 200,
  "amp": true,
  "nsamples": 128,
  "low_gpu_mem_usage": false,
  "to_quant_block_names": "language_model.model.layers",
  "enable_norm_bias_tuning": false,
  "dataset": "NeelNanda/pile-10k",
  "autoround_version": "0.4.6",
  "quant_method": "awq",
  "zero_point": false,
  "version": "gemm",
  "modules_to_not_convert": [
    "vision_tower.transformer.layers",
    "multi_modal_projector.patch_merger.merging_layer",
    "multi_modal_projector.linear_1",
    "multi_modal_projector.linear_2",
    "language_model.lm_head"
  ]
}