MiniMax-H3-OrbitQuant-W4A4 / scripts /decode_precision_policy.py
WaveCut's picture
Publish validated MiniMax H3 OrbitQuant W4 release
fa2d87b
Raw
History Blame Contribute Delete
263 Bytes
from __future__ import annotations
def torch_dtype_name(precision: str) -> str | None:
if precision in {"orbitquant", "fp32"}:
return None
if precision == "bf16":
return "bfloat16"
raise ValueError(f"unknown precision: {precision}")