[BugFix] fix real_bsz in ep (#3366)

* Your commit message here

* fix ep

* delete cuda_graph

fastdeploy/worker/worker_process.py

@@ -250,6 +250,7 @@ class PaddleDisWorkerProc:
         while True:
             self.worker_healthy_live_signal.value[self.local_rank % self.max_chips_per_node] = int(time.time())
 
+            num_running_requests = 0
             if self.fd_config.parallel_config.tensor_parallel_rank == 0 and self.task_queue.num_tasks() > 0:
                 tasks, read_finish = self.task_queue.get_tasks()
 
@@ -276,6 +277,7 @@ 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: