fp4-fp8-for-torch-mps 1.0.1

FP8 and FP4 sub-byte dtype support for PyTorch MPS on Apple Silicon via Metal shaders

FP8/FP4 support for PyTorch MPS on Apple Silicon.

Installing this lets ComfyUI use FP8 and FP4 weight checkpoints on MPS automatically.

Fast paths use Metal kernels. Fallback paths round-trip through CPU Torch for exact PyTorch-compatible bytes.

pip install fp4-fp8-for-torch-mps

Decode backend

FP8 e4m3fn and FP4 decode can use lut, lut4, arithmetic, arithmetic4, or auto. Default: lut.

import fp4_fp8_for_torch_mps as mps_fp
mps_fp.set_decode_backend("lut")

Or:

FP4_FP8_MPS_DECODE_BACKEND=lut