transformer/quantization_config.json

{
  "bits": 4,
  "act_bits": 4,
  "data_type": "mx_fp",
  "act_data_type": "mx_fp",
  "group_size": 32,
  "act_group_size": 32,
  "sym": true,
  "act_sym": true,
  "act_dynamic": true,
  "enable_quanted_input": false,
  "static_attention_granularity": "tensor",
  "static_kv_granularity": "tensor",
  "autoround_version": "0.15.0",
  "block_name_to_quantize": "transformer_blocks,single_transformer_blocks",
  "quant_method": "auto-round",
  "packing_format": "auto_round:llm_compressor"
}