update openai_api: support stop words for streaming chat

openai_api.py

@@ -483,20 +483,19 @@ async def predict(
     current_length = 0
     stop_words_ids = [tokenizer.encode(s)
                       for s in stop_words] if stop_words else None
-    if stop_words:
+
-        # TODO: It's a little bit tricky to trim stop words in the stream mode.
+    delay_token_num = max([len(x) for x in stop_words])
-        raise HTTPException(
-            status_code=400,
-            detail=
-            'Invalid request: custom stop words are not yet supported for stream mode.',
-        )
     response_generator = model.chat_stream(tokenizer,
                                            query,
                                            history=history,
                                            stop_words_ids=stop_words_ids,
                                            system=system,
                                            **gen_kwargs)
-    for new_response in response_generator:
+    for _new_response in response_generator:
+        if len(_new_response) <= delay_token_num:
+            continue 
+        new_response = _new_response[:-delay_token_num]
+
         if len(new_response) == current_length:
             continue
 
@@ -510,6 +509,18 @@ async def predict(
                                        object='chat.completion.chunk')
         yield '{}'.format(_dump_json(chunk, exclude_unset=True))
     
+    if current_length != len(_new_response):
+        # Determine whether to print the delay tokens
+        delayed_text = _new_response[current_length:]
+        new_text = trim_stop_words(delayed_text, stop_words)
+        if len(new_text) > 0:
+            choice_data = ChatCompletionResponseStreamChoice(
+                index=0, delta=DeltaMessage(content=new_text), finish_reason=None)
+            chunk = ChatCompletionResponse(model=model_id,
+                                        choices=[choice_data],
+                                        object='chat.completion.chunk')
+            yield '{}'.format(_dump_json(chunk, exclude_unset=True))
+
     choice_data = ChatCompletionResponseStreamChoice(index=0,
                                                      delta=DeltaMessage(),
                                                      finish_reason='stop')