diff --git a/fastdeploy/worker/gpu_model_runner.py b/fastdeploy/worker/gpu_model_runner.py
index ef6e4a200..50fc205b8 100644
--- a/fastdeploy/worker/gpu_model_runner.py
+++ b/fastdeploy/worker/gpu_model_runner.py
@@ -256,7 +256,9 @@ class GPUModelRunner(ModelRunnerBase):
                     )
 
                 input_ids = request.prompt_token_ids + request.output_token_ids
-                logger.debug(f"Handle prefill request {request} at idx {idx} prefill_start_index {prefill_start_index} prefill_end_index {prefill_end_index} need_prefilled_token_num {len(input_ids)}")
+                logger.debug(
+                    f"Handle prefill request {request} at idx {idx} prefill_start_index {prefill_start_index} prefill_end_index {prefill_end_index} need_prefilled_token_num {len(input_ids)}"
+                )
                 self.share_inputs["input_ids"][idx : idx + 1, :length] = np.array(
                     input_ids[prefill_start_index:prefill_end_index]
                 )
@@ -617,7 +619,9 @@ class GPUModelRunner(ModelRunnerBase):
         self.share_inputs["max_length"] = paddle.full(
             [max_num_seqs, 1], self.model_config.max_model_len, dtype="int64"
         )
-        self.seq_lens_this_time_buffer = paddle.full(max_num_seqs, 0, dtype="int32")
+        self.seq_lens_this_time_buffer = paddle.full([max_num_seqs, 1], 0, dtype="int32")
+        if self.fd_config.parallel_config.enable_expert_parallel:
+            self.share_inputs["seq_lens_this_time"] = paddle.full([max_num_seqs, 1], 0, dtype="int32")
         self.share_inputs["seq_lens_encoder"] = paddle.full([max_num_seqs, 1], 0, dtype="int32")
         self.share_inputs["seq_lens_decoder"] = paddle.full([max_num_seqs, 1], 0, dtype="int32")
         self.share_inputs["step_seq_lens_encoder"] = paddle.full([max_num_seqs, 1], 0, dtype="int32")
diff --git a/fastdeploy/worker/worker_process.py b/fastdeploy/worker/worker_process.py
index fc5026bdb..71ed1ff66 100644
--- a/fastdeploy/worker/worker_process.py
+++ b/fastdeploy/worker/worker_process.py
@@ -244,6 +244,7 @@ class PaddleDisWorkerProc:
         Tmp loop function for ep utill DP is supported
         """
         while True:
+            num_running_requests = 0
             self.worker_healthy_live_signal.value[self.local_rank % self.max_chips_per_node] = int(time.time())
 
             if self.fd_config.parallel_config.tensor_parallel_rank == 0 and self.task_queue.num_tasks() > 0:
@@ -272,6 +273,7 @@ class PaddleDisWorkerProc:
         self.nnode = int((self.parallel_config.tensor_parallel_size + 7) // 8)
         mp_num_per_node = self.parallel_config.tensor_parallel_size // self.nnode
         req_ids = []
+        num_running_requests = 0
         while True:
             if self.local_rank == 0:
                 if self.model_weights_status.value[0] != 0: