fix request_output sampling_params (#3154) (#3464)

Author: ckl117

fastdeploy/engine/engine.py

@@ -734,10 +734,6 @@ def insert_tasks(self, tasks, current_id=-1, allocated=False):
         """
         Insert tasks to engine.
         """
-        for task in tasks:
-            start_span_request("DEQUEUE", task, trace.SpanKind.CONSUMER)
-            if task.sampling_params.bad_words is not None:
-                task.sampling_params.update_from_tokenizer(self.data_processor.tokenizer)
         # TODO 返回至 scheduler
         if allocated:
             current_tasks = []
@@ -764,6 +760,11 @@ def insert_tasks(self, tasks, current_id=-1, allocated=False):
             self.engine_worker_queue.put_tasks((current_tasks, self.resource_manager.real_bsz))
             return True
 
+        for task in tasks:
+            start_span_request("DEQUEUE", task, trace.SpanKind.CONSUMER)
+            if task.sampling_params.bad_words is not None:
+                task.sampling_params.update_from_tokenizer(self.data_processor.tokenizer)
+
         self.resource_manager.check_and_free_block_tables()
 
         if not isinstance(tasks, list):