[MoE Refactor] MXFP4 Cutlass Experts to MK (#34542)

Signed-off-by: Yongye Zhu <zyy1102000@gmail.com>
This commit is contained in:
Yongye Zhu
2026-02-25 17:32:39 -08:00
committed by GitHub
parent cbf8f7028c
commit 1976356ee6
19 changed files with 454 additions and 169 deletions
@@ -19,6 +19,7 @@ if TYPE_CHECKING:
FP8_DTYPE = current_platform.fp8_dtype()
FP4_DTYPE = torch.uint8
MXFP_SCALE_DTYPE = torch.uint8
def get_fp8_min_max() -> tuple[float, float]:
@@ -151,6 +152,18 @@ kFp8Static128BlockSym = QuantKey(FP8_DTYPE, kStatic128BlockScale, symmetric=True
kDynamic64Scale = ScaleDesc(torch.float32, False, GroupShape(1, 64))
kFp8Dynamic64Sym = QuantKey(FP8_DTYPE, kDynamic64Scale, symmetric=True)
# TODO (zyongye): Convert all the torch.dtype to scale_dtype
# Changing that requires changing torch compile fused AR+Quant Quant key
# to avoid assertion error
kMxfp4DynamicGroupScale = ScaleDesc(MXFP_SCALE_DTYPE, False, GroupShape(1, 32))
kMxfp4Dynamic = QuantKey(FP4_DTYPE, scale=kMxfp4DynamicGroupScale, symmetric=True)
kMxfp8DynamicGroupScale = ScaleDesc(MXFP_SCALE_DTYPE, False, GroupShape(1, 32))
kMxfp8Dynamic = QuantKey(FP8_DTYPE, scale=kMxfp8DynamicGroupScale, symmetric=True)
kMxfp4StaticGroupScale = ScaleDesc(MXFP_SCALE_DTYPE, True, GroupShape(1, 32))
kMxfp4Static = QuantKey(FP4_DTYPE, scale=kMxfp4StaticGroupScale, symmetric=True)
# Normalize the group_shape to the full extent for any dims that are -1
def _normalize_quant_group_shape(x: torch.Tensor, group_shape: GroupShape):