diff --git a/fastdeploy/model_executor/models/glm4_moe.py b/fastdeploy/model_executor/models/glm4_moe.py
index 3f45e9df61..b32ebb2ced 100644
--- a/fastdeploy/model_executor/models/glm4_moe.py
+++ b/fastdeploy/model_executor/models/glm4_moe.py
@@ -26,6 +26,7 @@ from paddleformers.transformers import PretrainedModel
 from paddleformers.utils.log import logger
 
 from fastdeploy.config import FDConfig
+from fastdeploy.distributed.communication import tensor_model_parallel_all_reduce
 from fastdeploy.model_executor.forward_meta import ForwardMeta
 from fastdeploy.model_executor.graph_optimization.decorator import (
     support_graph_optimization,
@@ -160,8 +161,16 @@ class Glm4Moe(nn.Layer):
             default_initializer=paddle.nn.initializer.Constant(0),
         )
 
+        # In pure-TP mode (tp>1, ep=1) both branches return partial sums, so we
+        # defer the all-reduce to after combining them — saving one collective.
+        # In all other modes (EP, EP+attn-TP, no parallelism) each branch handles
+        # its own reduction internally (reduce_results default=True), so we must
+        # NOT add an extra all-reduce here.
+        self.merge_ffn_tp = self.use_tp and not self.use_ep
+
         self.experts = FusedMoE(
             fd_config,
+            reduce_results=not self.merge_ffn_tp,
             renormalize=self.norm_topk_prob,
             moe_intermediate_size=fd_config.model_config.moe_intermediate_size,
             num_experts=fd_config.model_config.n_routed_experts,
@@ -182,14 +191,16 @@ class Glm4Moe(nn.Layer):
                 intermediate_size=shared_experts_intermediate_size,
                 layer_id=layer_id,
                 prefix=f"{prefix}.shared_experts",
+                reduce_results=not self.merge_ffn_tp,
             )
 
     def forward(self, x, forward_meta: ForwardMeta = None):
         out = self.experts(x, self.gate, forward_meta)
         if self.n_shared_experts > 0:
-            shared_experts_out = self.shared_experts(x)
-            out = out + shared_experts_out
-
+            out = out + self.shared_experts(x)
+        if self.merge_ffn_tp:
+            # Both branches produced partial sums; combine first, then single all-reduce.
+            out = tensor_model_parallel_all_reduce(out, self.tp_group)
         return out
 
 
diff --git a/tests/e2e/4cards_cases/test_GLM_45_AIR_mtp_tp4.py b/tests/e2e/4cards_cases/test_GLM_45_AIR_mtp_tp4.py
index 4024b57613..b3bd6d9b45 100644
--- a/tests/e2e/4cards_cases/test_GLM_45_AIR_mtp_tp4.py
+++ b/tests/e2e/4cards_cases/test_GLM_45_AIR_mtp_tp4.py
@@ -185,7 +185,7 @@ def test_lm_head_fp32(api_url, headers, consistent_payload):
     # 校验返回内容与概率信息
     assert (
         resp_json["choices"][0]["message"]["content"]
-        == "\n<think>这个问题是关于牛顿的三大运动定律的。牛顿的三大运动定律是经典"
+        == "\n<think>我需要回答牛顿的三大运动定律是什么。牛顿的三大运动定律是经典"
     ), f"The response content is not as expected {resp_json['choices'][0]['message']['content']}."
 
 
diff --git a/tests/e2e/utils/rollout_routing_replay_test_utils.py b/tests/e2e/utils/rollout_routing_replay_test_utils.py
index 8a01b50240..4186a71649 100644
--- a/tests/e2e/utils/rollout_routing_replay_test_utils.py
+++ b/tests/e2e/utils/rollout_routing_replay_test_utils.py
@@ -157,10 +157,10 @@ def check_routing_replay_chat_completion(openai_client, moe_layer_num: int, mode
     model_path = os.getenv("MODEL_PATH")
     if model_path:
         baseline_path = os.path.join(
-            model_path, f"R3_BaseLine_dev_uint8_0312/routing_replay_output_baseline_{model_name}"
+            model_path, f"R3_BaseLine_dev_uint8_0402/routing_replay_output_baseline_{model_name}"
         )
     else:
-        baseline_path = f"./R3_BaseLine_dev_uint8_0312/routing_replay_output_baseline_{model_name}"
+        baseline_path = f"./R3_BaseLine_dev_uint8_0402/routing_replay_output_baseline_{model_name}"
     stream_baseline_path = os.path.join(baseline_path, "r3_chat_completion_stream")
 
     nonstream_baseline_path = os.path.join(baseline_path, "r3_chat_completion_nonstream")