[BugFix] fix config bugs (#4370)

* Update expert_service.py

* Update common_engine.py

* Update expert_service.py

* Update expert_service.py

* Update expert_service.py

---------

Co-authored-by: Jiang-Jia-Jun <[email protected]>

Author: ltd0924

fastdeploy/engine/common_engine.py

@@ -68,12 +68,13 @@ def __init__(self, cfg, start_queue=True):
             cfg (Config): Config object containing all the configuration parameters.
         """
         self.cfg = cfg
-        if isinstance(self.cfg.cache_config.cache_queue_port, str):
-            self.cfg.cache_config.cache_queue_port = self.cfg.cache_config.cache_queue_port.split(",")
-        if isinstance(self.cfg.cache_config.cache_queue_port, list):
-            self.cfg.cache_config.cache_queue_port = int(
-                self.cfg.cache_config.cache_queue_port[self.cfg.parallel_config.local_data_parallel_id]
-            )
+        if cfg.scheduler_config.splitwise_role != "mixed" or cfg.cache_config.enable_prefix_caching:
+            if isinstance(self.cfg.cache_config.cache_queue_port, str):
+                self.cfg.cache_config.cache_queue_port = self.cfg.cache_config.cache_queue_port.split(",")
+            if isinstance(self.cfg.cache_config.cache_queue_port, list):
+                self.cfg.cache_config.cache_queue_port = int(
+                    self.cfg.cache_config.cache_queue_port[self.cfg.parallel_config.local_data_parallel_id]
+                )
 
         if self.cfg.parallel_config.enable_expert_parallel:
             self.llm_logger = get_logger(

fastdeploy/engine/expert_service.py

@@ -103,29 +103,6 @@ def start(
 
         llm_logger.info(f"start expert service {local_data_parallel_id}")
 
-        if self.cfg.scheduler_config.splitwise_role != "mixed" or self.cfg.cache_config.enable_prefix_caching:
-            if self.do_profile:
-                get_profile_block_num = np.zeros([1], dtype=np.int32)
-                while True:
-                    try:
-                        self.get_profile_block_num_signal = IPCSignal(
-                            name="get_profile_block_num",
-                            array=get_profile_block_num,
-                            dtype=np.int32,
-                            suffix=int(self.cfg.engine_worker_queue_port[0]),
-                            create=False,
-                        )
-                        break
-                    except:
-                        time.sleep(1)
-                self.reset_kvcache_blocks()
-            ipc_signal_suffix_cache = self.cfg.parallel_config.engine_worker_queue_port[local_data_parallel_id]
-            self.cache_manager_processes = self.engine.start_cache_service(
-                self.cfg.local_device_ids, ipc_signal_suffix_cache
-            )
-            if self.cfg.scheduler_config.splitwise_role != "mixed":
-                self.engine.split_mode_get_tasks()
-
         if self.cfg.scheduler_config.name == "splitwise":
             self.cfg.init_cache_info()
             role = self.cfg.scheduler_config.splitwise_role
@@ -155,6 +132,29 @@ def start(
             )
             self.launched_expert_service_signal.value[local_rank] = 1
 
+        if self.cfg.scheduler_config.splitwise_role != "mixed" or self.cfg.cache_config.enable_prefix_caching:
+            if self.do_profile:
+                get_profile_block_num = np.zeros([1], dtype=np.int32)
+                while True:
+                    try:
+                        self.get_profile_block_num_signal = IPCSignal(
+                            name="get_profile_block_num",
+                            array=get_profile_block_num,
+                            dtype=np.int32,
+                            suffix=int(self.cfg.parallel_config.engine_worker_queue_port[0]),
+                            create=False,
+                        )
+                        break
+                    except:
+                        time.sleep(1)
+                self.reset_kvcache_blocks()
+            ipc_signal_suffix_cache = self.cfg.parallel_config.engine_worker_queue_port[local_data_parallel_id]
+            self.cache_manager_processes = self.engine.start_cache_service(
+                self.cfg.local_device_ids,
+                ipc_signal_suffix_cache,
+                create_cache_tensor=(self.cfg.scheduler_config.splitwise_role != "mixed"),
+            )
+
         console_logger.info(
             f"Worker processes(rank {local_rank}) are launched with {time.time() - start_time} seconds."
         )