[Feature] support bad_words (#3055)

* support bad_words

* support online infer bad_words

* update

* add CI test

* update

* update

* update

---------

Co-authored-by: Yuanle Liu <yuanlehome@163.com>

fastdeploy/engine/engine.py

@@ -491,6 +491,7 @@ class LLMEngine:
         request = Request.from_dict(task)
         llm_logger.info(f"Receive request {request}")
         if sampling_params is not None:
+            sampling_params.update_from_tokenizer(self.data_processor.tokenizer)
             request.sampling_params = sampling_params
         request.preprocess_start_time = time.time()
 
@@ -747,6 +748,8 @@ class LLMEngine:
         """
         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 = []