Merge branch 'develop' into fastdeploy_v100

mattheliu · web-flow · commit 8150b9cbdf81 · 2026-02-09T13:21:59.000+08:00
diff --git a/custom_ops/gpu_ops/noaux_tc_redundant.cu b/custom_ops/gpu_ops/noaux_tc_redundant.cu
@@ -61,7 +61,8 @@ std::vector<paddle::Tensor> NoauxTcRedundant(
       redundant_ep_rank_num_plus_one,
       stream);
 
-  return {scores, topk_values, topk_indices};
+  // Return 4 values to match PD_BUILD_STATIC_OP Outputs definition
+  return {scores, topk_values, topk_indices, tokens_per_expert_stats_list};
 }
 
 std::vector<paddle::DataType> NoauxTcRedundantInferDtype(
diff --git a/fastdeploy/model_executor/layers/moe/moe.py b/fastdeploy/model_executor/layers/moe/moe.py
@@ -104,7 +104,9 @@ def get_moe_scores(
             routed_scaling_factor,
         )
     else:
-        scores, topk_values, topk_idx = noaux_tc_redundant(
+        # noaux_tc_redundant returns 4 values: scores, topk_values, topk_idx,
+        # and tokens_per_expert_stats_list_out (inplace updated)
+        scores, topk_values, topk_idx, _ = noaux_tc_redundant(
             scores,
             scores_with_bias,
             expert_id_to_ep_rank_array,

Original file line number	Diff line number	Diff line change
`@@ -61,7 +61,8 @@ std::vector<paddle::Tensor> NoauxTcRedundant(`
`61`	`61`	`redundant_ep_rank_num_plus_one,`
`62`	`62`	`stream);`
`63`	`63`
`64`		`- return {scores, topk_values, topk_indices};`
	`64`	`+ // Return 4 values to match PD_BUILD_STATIC_OP Outputs definition`
	`65`	`+ return {scores, topk_values, topk_indices, tokens_per_expert_stats_list};`
`65`	`66`	`}`
`66`	`67`
`67`	`68`	`std::vector<paddle::DataType> NoauxTcRedundantInferDtype(`