Bugfix: Cutlass FP8 FusedMoE bad scaling factors (#27255)

Signed-off-by: Amir Klein <203507526+amirkl94@users.noreply.github.com>
Co-authored-by: Tyler Michael Smith <tyler@neuralmagic.com>
Co-authored-by: Michael Goin <mgoin64@gmail.com>
This commit is contained in:
amirkl94
2025-11-05 06:06:06 -05:00
committed by GitHub
co-authored by Tyler Michael Smith Michael Goin
parent b57789b62b
commit 6b7a81185d
4 changed files with 40 additions and 14 deletions
@@ -567,9 +567,13 @@ class ModelOptFp8MoEMethod(FusedMoEMethodBase):
return fp8_w8a8_moe_quant_config(
w1_scale=layer.w13_weight_scale,
g1_alphas=(layer.w13_weight_scale * layer.w13_input_scale).squeeze(),
w2_scale=layer.w2_weight_scale,
g2_alphas=(layer.w2_weight_scale * layer.w2_input_scale).squeeze(),
a1_scale=layer.w13_input_scale,
a1_gscale=layer.w13_input_scale,
a2_scale=layer.w2_input_scale,
a2_gscale=1.0 / layer.w2_input_scale,
per_act_token_quant=False,
)
@@ -1138,8 +1142,8 @@ class ModelOptNvFp4FusedMoE(FusedMoEMethodBase):
moe: FusedMoEConfig,
layer: torch.nn.Module,
) -> None:
from vllm.model_executor.layers.quantization.utils.nvfp4_moe_support import ( # noqa: E501
detect_nvfp4_moe_support,
from vllm.model_executor.layers.quantization.utils.nvfp4_moe_support import (
detect_nvfp4_moe_support, # noqa: E501
)
super().__init__(moe)