[BugFix] fix real_bsz in ep (#3366)

* Your commit message here

* fix ep

* delete cuda_graph

Author: lizexu123

fastdeploy/model_executor/models/ernie4_5_moe.py

@@ -450,7 +450,7 @@ def empty_input_forward(self):
             self.fd_config.model_config.moe_layer_start_index,
             self.fd_config.model_config.num_hidden_layers,
         ):
-            self.ernie.layers[i].mlp.expert(fake_hidden_states)
+            self.ernie.layers[i].mlp.experts(fake_hidden_states, self.ernie.layers[i].mlp.gate)
 
     def forward(
         self,

fastdeploy/worker/gpu_model_runner.py

@@ -636,7 +636,9 @@ def _init_share_inputs(self, max_num_seqs: int):
         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")

fastdeploy/worker/worker_process.py

@@ -250,6 +250,7 @@ def event_loop_ep(self) -> None:
         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 @@ def event_loop_normal(self) -> None:
         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: