[MoE refactor] refactor GPTQMarlinMoEMethod with MK (#37990)

Signed-off-by: Kunshang Ji <kunshang.ji@intel.com>
Signed-off-by: Robert Shaw <robertgshaw2@gmail.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
Co-authored-by: Robert Shaw <robertgshaw2@gmail.com>
This commit is contained in:
Kunshang Ji
2026-04-23 05:21:47 +00:00
committed by GitHub
co-authored by Robert Shaw Robert Shaw
parent 342c58bc54
commit e4ee48da2d
4 changed files with 577 additions and 228 deletions
@@ -20,6 +20,8 @@ if TYPE_CHECKING:
FP8_DTYPE = current_platform.fp8_dtype()
FP4_DTYPE = torch.uint8
MXFP_SCALE_DTYPE = torch.uint8
INT4_DTYPE = scalar_types.uint4b8
INT8_DTYPE = scalar_types.uint8b128
def get_fp8_min_max() -> tuple[float, float]:
@@ -170,6 +172,12 @@ kMxfp8Dynamic = QuantKey(FP8_DTYPE, scale=kMxfp8DynamicGroupScale, symmetric=Tru
kMxfp4StaticGroupScale = ScaleDesc(MXFP_SCALE_DTYPE, True, GroupShape(1, 32))
kMxfp4Static = QuantKey(FP4_DTYPE, scale=kMxfp4StaticGroupScale, symmetric=True)
# TODO: convert this to use SCALAR_TYPE. This is not right.
kInt4StaticGroupScale = ScaleDesc(torch.float16, True, GroupShape(1, -1))
kInt4Static = QuantKey(INT4_DTYPE, scale=kInt4StaticGroupScale, symmetric=True)
kInt8StaticGroupScale = ScaleDesc(torch.float16, True, GroupShape(1, -1))
kInt8Static = QuantKey(INT8_DTYPE, scale=kInt8StaticGroupScale, symmetric=True)
kInt8StaticChannelSym = QuantKey(torch.int8, kStaticChannelScale, symmetric=True)
kInt8DynamicTokenSym = QuantKey(torch.int8, kDynamicTokenScale, symmetric=True)