quantization_config.json

{
  "bits": 4,
  "act_bits": 4,
  "data_type": "nv_fp",
  "act_data_type": "nv_fp4_with_static_gs",
  "group_size": 16,
  "act_group_size": 16,
  "sym": true,
  "act_sym": true,
  "act_dynamic": true,
  "seqlen": 1024,
  "autoround_version": "0.12.3",
  "quant_method": "auto-round",
  "packing_format": "auto_round:llm_compressor"
}