fix ut

Author: zoooo0820

fastdeploy/model_executor/layers/moe/moe.py

@@ -86,7 +86,7 @@ def get_moe_scores(
     expert_in_rank_num_list: paddle.Tensor = None,
     tokens_per_expert_stats_list: paddle.Tensor = None,
     redundant_ep_rank_num_plus_one: int = 1,
-    topk_reduce_func: Callable = None,
+    topk_reduce_func: Callable = lambda x: x.sum(axis=-1, keepdim=True) + 1e-20,
 ) -> paddle.Tensor:
     """
     compute moe scores using e_score_correction_bias.
@@ -129,7 +129,7 @@ def get_moe_scores(
             redundant_ep_rank_num_plus_one,
         )
     if envs.FD_USE_PHI_MOE_TOPK:
-        if topk_reduce_func is not None and original_renormalize:
+        if original_renormalize:
             topk_values = topk_values / topk_reduce_func(topk_values)
 
             if original_routed_scaling_factor != 1.0:
@@ -163,7 +163,8 @@ def __init__(
         with_bias: bool = False,
         activation="swiglu",
         model_format: Optional[str] = None,
-        topk_reduce_func: Callable = None,  # only used when FD_USE_PHI_MOE_TOPK=1
+        topk_reduce_func: Callable = lambda x: x.sum(axis=-1, keepdim=True)
+        + 1e-20,  # only used when FD_USE_PHI_MOE_TOPK=1, default is same as noaux_tc kernel
     ):
         """
         Initialize the Moe layer with given parameters.

tests/operators/test_noaux_tc_redundant.py

@@ -1,10 +1,22 @@
+# Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import unittest
+from unittest import mock
 
 import paddle
 
-from fastdeploy.model_executor.layers.moe.fused_moe_deepgemm_backend import (
-    moe_topk_select,
-)
 from fastdeploy.model_executor.layers.moe.moe import get_moe_scores
 
 
@@ -135,15 +147,17 @@ def test_group_topk_using_phi_topk(self):
                     e_score_correction_bias=e_score_correction_bias,
                 )
 
-                topk_values, topk_idx = moe_topk_select(
-                    gating_output=gating_output,
-                    n_group=n_group,
-                    topk_group=topk_group,
-                    top_k=top_k,
-                    routed_scaling_factor=routed_scaling_factor,
-                    e_score_correction_bias=e_score_correction_bias,
-                    renormalize=renormalize,
-                )
+                with mock.patch.dict("os.environ", {"FD_USE_PHI_MOE_TOPK": "1"}):
+                    new_score, topk_values, topk_idx = get_moe_scores(
+                        gating_output=gating_output,
+                        n_group=n_group,
+                        topk_group=topk_group,
+                        top_k=top_k,
+                        routed_scaling_factor=routed_scaling_factor,
+                        e_score_correction_bias=e_score_correction_bias,
+                        renormalize=renormalize,
+                        topk_reduce_func=lambda x: x.sum(axis=-1, keepdim=True) + 1e-20,
+                    )
 
                 equal_topk_value = paddle.allclose(topk_values, ref_topk_values, atol=1e-03, rtol=1e-03).item()
                 equal_topk_ids = paddle.allclose(