flashinfer.quantization.NVFP44Over6Config

class flashinfer.quantization.NVFP44Over6Config(e4m3_max: int = 448, err_mode: NVFP44Over6ErrMode | str = NVFP44Over6ErrMode.MAE, err_use_fast_math: bool = False)

NVFP4 4over6 configuration shared by Python drivers and kernels.

Parameters:

  • e4m3_max (int) – Upper bound of the E4M3 block scale, either 448 (full range) or 256. Also pins the per-tensor global scale to 1 / (e4m3_max * 6).

  • err_mode (NVFP44Over6ErrMode or str) – Error metric used to choose between the two scale candidates. A string ("MAE" / "MSE") is normalized to the enum member.

  • err_use_fast_math (bool) – Evaluate the candidate error in fp16 rather than exactly.

__init__(e4m3_max: int = 448, err_mode: NVFP44Over6ErrMode | str = NVFP44Over6ErrMode.MAE, err_use_fast_math: bool = False) → None

Methods

__init__([e4m3_max, err_mode, err_use_fast_math])

Attributes

e4m3_max

err_mode

err_mode_name

err_use_fast_math