diff --git a/vllm/model_executor/layers/quantization/fp8.py b/vllm/model_executor/layers/quantization/fp8.py
index 65e0b7062153..d2616da84a00 100644
--- a/vllm/model_executor/layers/quantization/fp8.py
+++ b/vllm/model_executor/layers/quantization/fp8.py
@@ -757,10 +757,9 @@ class Fp8MoEMethod(FusedMoEMethodBase):
                 layer.w2_weight = torch.nn.Parameter(shuffled_w2,
                                                      requires_grad=False)
 
-            # DeepGemm scales need to be transposed and aligned.  We try to do
+            # DeepGemm scales need to be transposed and aligned. We try to do
             # it ahead of time for performance reasons.
             if self.allow_deep_gemm and not is_deep_gemm_e8m0_used():
-                # Lazy import to avoid CUDA initialization problems.
                 if _is_col_major(layer.w13_weight_scale_inv):
                     layer.w13_weight_scale_inv = \
                         get_col_major_tma_aligned_tensor(layer.w13_weight_scale_inv).contiguous()