models :
  name : qwen2.5_fp4
  quant:
  - methods : fp4
    regex : layers
    type : fp4linear
    kwargs:
      block_size : 16