[Performance][Hopper] Avoid M dim padding to 4x for most cases (due to cuda...

[Performance][Hopper] Avoid M dim padding to 4x for most cases (due to cuda graphs paddings) (#28492)
Signed-off-by: Alexander Matveev <amatveev@redhat.com>

vllm/model_executor/layers/quantization/utils/fp8_utils.py

@@ -115,6 +115,9 @@ def _padded_cutlass(
         dim if dim % pad_multiple == 0 else dim + pad_multiple - (dim % pad_multiple)
     )
 
+    has_pad = padded > dim
+
+    if has_pad:
         padded_shape = [padded, *qx.shape[1:]]
         padded_qx = torch.zeros(padded_shape, device=qx.device, dtype=qx.dtype)
         padded_qx[0 : qx.shape[0], ...].copy_(qx)
@@ -129,6 +132,10 @@ def _padded_cutlass(
             padded_qx, weight, padded_x_scale, weight_scale, block_size, output_dtype
         )
         return output[0 : qx.shape[0], ...]
+    else:
+        return cutlass_scaled_mm(
+            qx, weight, x_scale, weight_scale, block_size, output_dtype
+        )
 
 
 def _padded_cutlass_fake(