diff --git a/python/sglang/srt/hardware_backend/npu/moe/topk.py b/python/sglang/srt/hardware_backend/npu/moe/topk.py
index 9aef68476..d59498038 100644
--- a/python/sglang/srt/hardware_backend/npu/moe/topk.py
+++ b/python/sglang/srt/hardware_backend/npu/moe/topk.py
@@ -74,7 +74,8 @@ def fused_topk_npu(
         topk_weights = scores.gather(1, topk_ids)
         if renormalize:
             topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)
-            topk_weights = _apply_routed_scaling_after_renorm(topk_weights, topk_config)
+        else:
+            topk_weights = topk_weights * topk_config.routed_scaling_factor
         topk_weights = topk_weights.to(torch.float32)
 
     # Support grouped top-k or correction bias or sigmoid or routed_scaling_factor
diff --git a/python/sglang/srt/models/llada2.py b/python/sglang/srt/models/llada2.py
index 94aec91b8..5fa3f64d8 100644
--- a/python/sglang/srt/models/llada2.py
+++ b/python/sglang/srt/models/llada2.py
@@ -265,6 +265,7 @@ class LLaDA2MoeSparseMoeBlock(nn.Module):
             correction_bias=self.correction_bias,
             scoring_func=self.score_function,
             routed_scaling_factor=self.routed_scaling_factor,
+            apply_routed_scaling_factor_on_output=_is_npu,
         )
 
         self.experts = get_moe_impl_class(quant_config)(
diff --git a/test/registered/ascend/basic_function/dllm/test_npu_llada2_mini.py b/test/registered/ascend/basic_function/dllm/test_npu_llada2_mini.py
index 467b73dfc..b8216893a 100644
--- a/test/registered/ascend/basic_function/dllm/test_npu_llada2_mini.py
+++ b/test/registered/ascend/basic_function/dllm/test_npu_llada2_mini.py
@@ -29,6 +29,7 @@ class TestLLaDA2Mini(GSM8KAscendMixin, CustomTestCase):
         "ascend",
         "--dllm-algorithm",
         "LowConfidence",  # TODO: Add dLLM configurations
+        "--no-dllm-fdfo",  # FDFO (PR #27551) halves single-batch speed on NPU; use sync mode
     ]
     env = {
         **os.environ,