fix `finish_reason` (#816)

lmdeploy/serve/async_engine.py

@@ -204,7 +204,7 @@ class AsyncEngine:
         if do_preprocess:
             prompt = self.model.messages2prompt(prompt, sequence_start)
         input_ids = self.tokenizer.encode(prompt, add_bos=sequence_start)
-        finish_reason = 'stop' if stop else None
+        finish_reason = None
         if self.id2step[str(session_id)] + len(
                 input_ids) + request_output_len >= self.tm_model.session_len:
             finish_reason = 'length'
@@ -247,9 +247,12 @@ class AsyncEngine:
                                  len(input_ids), tokens, finish_reason)
                     response_size = tokens
 
+                finish_reason = 'length' \
+                    if tokens >= request_output_len else 'stop'
                 # `response_size` might be note updated since
                 # ` if response.endswith('�')`
-                if response_size != tokens:
+                if response_size == tokens:
+                    response = ''  # avaid returning the last response twice
                 yield GenOut(response, self.id2step[str(session_id)],
                              len(input_ids), tokens, finish_reason)
                 # update step

lmdeploy/serve/openai/api_server.py

@@ -179,6 +179,7 @@ async def chat_completions_v1(request: ChatCompletionRequest,
             response_json = create_stream_response_json(
                 index=0,
                 text=res.response,
+                finish_reason=res.finish_reason,
             )
             yield f'data: {response_json}\n\n'
         yield 'data: [DONE]\n\n'
@@ -329,6 +330,7 @@ async def completions_v1(request: CompletionRequest,
                 response_json = create_stream_response_json(
                     index=0,
                     text=res.response,
+                    finish_reason=res.finish_reason,
                 )
                 yield f'data: {response_json}\n\n'
         yield 'data: [DONE]\n\n'