Fix topk inference performance reduce (#6474)

python/sglang/srt/layers/moe/topk.py

@@ -264,6 +264,8 @@ def biased_grouped_topk(
         # TODO merge into kernel for this branch
         topk_ids = topk_ids_logical_to_physical(topk_ids, expert_location_dispatch_info)
         # TODO will fuse this into kernel, thus use slow manual operation now
+        if num_token_non_padded is None:
+            return topk_weights, topk_ids
         torch.compile(
             _mask_topk_ids_padded_region, dynamic=True, backend=get_compiler_backend()
         )(topk_ids, num_token_non_padded)