Add NVFP4 MOE support for Deepseek V4. (#42209)

Signed-off-by: Shiyang Chen <shiychen@nvidia.com>
This commit is contained in:
sychen52
2026-05-22 07:21:51 -07:00
committed by GitHub
parent a377631d21
commit fb21d8b4f9
9 changed files with 217 additions and 17 deletions
@@ -258,6 +258,7 @@ class CompressedTensorsW4A4Nvfp4MoEMethod(CompressedTensorsMoEMethod):
w2_scale_2=layer.w2_weight_scale_2,
a13_scale=layer.w13_input_scale,
a2_scale=layer.w2_input_scale,
swiglu_limit=getattr(layer, "swiglu_limit", None),
)
def apply_monolithic(
@@ -1601,6 +1601,7 @@ class ModelOptNvFp4FusedMoE(FusedMoEMethodBase):
w2_scale_2=layer.w2_weight_scale_2,
a13_scale=layer.w13_input_scale,
a2_scale=layer.w2_input_scale,
swiglu_limit=getattr(layer, "swiglu_limit", None),
)
@property