tilelang.language.fp8¶

Functions¶

determine_fp8_type([fp8_format, device])

Select the correct FP8 dtype string for the current platform.

determine_torch_fp8_type([fp8_format, device])

Module Contents¶

tilelang.language.fp8.determine_fp8_type(fp8_format='e4m3', device=None)¶

Select the correct FP8 dtype string for the current platform. - CUDA defaults to FP8 E4M3FN / E5M2. - ROCm uses FNUZ except gfx950 (OCP), which prefers non-FNUZ when available.

Parameters:
  • fp8_format (Literal['e4m3', 'e5m2'])

  • device (torch.device | int | str | None)

Return type:

str

tilelang.language.fp8.determine_torch_fp8_type(fp8_format='e4m3', device=None)¶
Parameters:
  • fp8_format (Literal['e4m3', 'e5m2'])

  • device (torch.device | int | str | None)

Return type:

torch.dtype