quantization_config.json

{
  "bits": 8,
  "group_size": -1,
  "sym": true,
  "data_type": "fp8",
  "batch_size": 2,
  "iters": 0,
  "nsamples": 2,
  "act_bits": 8,
  "act_group_size": -1,
  "act_sym": true,
  "act_dynamic": false,
  "act_data_type": "fp8",
  "autoround_version": "0.6.0",
  "quant_method": "auto-round",
  "fmt": "e4m3",
  "activation_scheme": "dynamic"
}