quantization_config.json

{
  "bits": 4,
  "group_size": 16,
  "sym": true,
  "data_type": "nv_fp4",
  "seqlen": 512,
  "batch_size": 4,
  "iters": 20,
  "autoround_version": "0.6.0",
  "quant_method": "auto-round",
  "packing_format": "nv_fp",
  "scale_format": [
    "e8m0"
  ],
  "scale_calculation_mode": [
    "even"
  ]
}