Fix mismatch between padded_scales shape and reshape dimensions in modelopt quantization (#8766)

python/sglang/srt/layers/quantization/modelopt_quant.py

@@ -677,9 +677,9 @@ class ModelOptFp4LinearMethod(LinearMethodBase):
         padded_scales = padded_scales.permute((0, 1, 4, 3, 2, 5))
         padded_scales = padded_scales.contiguous().cuda()
         padded_scales = (
-            padded_scales.reshape(M, K)
+            padded_scales.reshape(M_padded, K_padded)
             if scale_ndim == 2
-            else padded_scales.reshape(B, M, K)
+            else padded_scales.reshape(B, M_padded, K_padded)
         )
         layer.weight_scale_interleaved = Parameter(padded_scales, requires_grad=False)
 
@@ -878,9 +878,9 @@ class ModelOptNvFp4FusedMoEMethod(FusedMoEMethodBase):
         swizzled_scale = padded_scale.permute((0, 1, 4, 3, 2, 5))
         swizzled_scale = swizzled_scale.contiguous().cuda()
         return (
-            swizzled_scale.reshape(M, K)
+            swizzled_scale.reshape(M_padded, K_padded)
             if scale_ndim == 2
-            else swizzled_scale.reshape(B, M, K)
+            else swizzled_scale.reshape(B, M_padded, K_padded)
         )
 
     def prepare_static_weights_for_kernel(