A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[Doc] Remove Useless Comments (#24687)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Wentao Ye committed
Sep 11, 2025 at 15:25 UTC
a892b259b4503526511d9ec17d5f8961e46969a6
1 file changed
+1
-2
vllm/model_executor/layers/quantization/fp8.py
+1
-2
@@ -757,10 +757,9 @@ class Fp8MoEMethod(FusedMoEMethodBase):
757
layer.w2_weight = torch.nn.Parameter(shuffled_w2,
758
requires_grad=False)
759
760
- # DeepGemm scales need to be transposed and aligned. We try to do
760
+ # DeepGemm scales need to be transposed and aligned. We try to do
761
# it ahead of time for performance reasons.
762
if self.allow_deep_gemm and not is_deep_gemm_e8m0_used():
763
- # Lazy import to avoid CUDA initialization problems.
763
if _is_col_major(layer.w13_weight_scale_inv):
764
layer.w13_weight_scale_inv = \
765
get_col_major_tma_aligned_tensor(layer.w13_weight_scale_inv).contiguous()