[BugFix] v1/completions add finish_reason

Author: memoryCoderC

fastdeploy/entrypoints/openai/serving_completion.py

@@ -224,6 +224,15 @@ async def completion_full_generator(
             if dealer is not None:
                 dealer.close()
 
+    def calc_finish_reason(self, max_tokens, token_num, output):
+        if max_tokens is None or token_num != max_tokens:
+            if self.engine_client.reasoning_parser == "ernie_x1" and output.get("finish_reason", "") == "tool_calls":
+                return "tool_calls"
+            else:
+                return "stop"
+        else:
+            return "length"
+
     async def completion_stream_generator(
         self,
         request: CompletionRequest,
@@ -324,19 +333,13 @@ async def completion_stream_generator(
                             logprobs=logprobs_res,
                         )
                     )
-                    if res["finished"]:
-                        if request.max_tokens is None or output_tokens[idx] + 1 != request.max_tokens:
-                            chunk.choices[0].finish_reason = "stop"
-                            if (
-                                self.engine_client.reasoning_parser == "ernie_x1"
-                                and output.get("finish_reason", "") == "tool_calls"
-                            ):
-                                chunk.choices[0].finish_reason = "tool_calls"
-                        else:
-                            chunk.choices[0].finish_reason = "length"
-
                     output_tokens[idx] += 1
 
+                    if res["finished"]:
+                        choices[-1].finish_reason = self.calc_finish_reason(
+                            request.max_tokens, output_tokens[idx], output
+                        )
+
                     if len(choices) == max_streaming_response_tokens or res["finished"]:
                         chunk = CompletionStreamResponse(
                             id=request_id,
@@ -421,6 +424,11 @@ def request_output_to_completion_response(
                 token_ids = output["token_ids"]
                 output_text = output["text"]
 
+            num_generated_tokens += final_res["output_token_ids"]
+            num_prompt_tokens += len(prompt_token_ids)
+
+            finish_reason = self.calc_finish_reason(request.max_tokens, final_res["output_token_ids"], output)
+
             choice_data = CompletionResponseChoice(
                 token_ids=token_ids,
                 index=len(choices),
@@ -430,14 +438,10 @@ def request_output_to_completion_response(
                 reasoning_content=output.get("reasoning_content"),
                 tool_calls=output.get("tool_call_content"),
                 logprobs=aggregated_logprobs,
-                finish_reason=None,
+                finish_reason=finish_reason,
             )
             choices.append(choice_data)
 
-            num_generated_tokens += final_res["output_token_ids"]
-
-            num_prompt_tokens += len(prompt_token_ids)
-
         usage = UsageInfo(
             prompt_tokens=num_prompt_tokens,
             completion_tokens=num_generated_tokens,

test/entrypoints/openai/test_serving_completion.py

@@ -0,0 +1,49 @@
+import unittest
+from unittest.mock import Mock
+
+from fastdeploy.entrypoints.openai.serving_completion import OpenAIServingCompletion
+
+
+class TestOpenAIServingCompletion(unittest.TestCase):
+
+    def test_calc_finish_reason_tool_calls(self):
+        # 创建一个模拟的engine_client，并设置reasoning_parser为"ernie_x1"
+        engine_client = Mock()
+        engine_client.reasoning_parser = "ernie_x1"
+        # 创建一个OpenAIServingCompletion实例
+        serving_completion = OpenAIServingCompletion(engine_client, "pid", "ips")
+        # 创建一个模拟的output，并设置finish_reason为"tool_calls"
+        output = {"finish_reason": "tool_calls"}
+        # 调用calc_finish_reason方法
+        result = serving_completion.calc_finish_reason(None, 100, output)
+        # 断言结果为"tool_calls"
+        assert result == "tool_calls"
+
+    def test_calc_finish_reason_stop(self):
+        # 创建一个模拟的engine_client，并设置reasoning_parser为"ernie_x1"
+        engine_client = Mock()
+        engine_client.reasoning_parser = "ernie_x1"
+        # 创建一个OpenAIServingCompletion实例
+        serving_completion = OpenAIServingCompletion(engine_client, "pid", "ips")
+        # 创建一个模拟的output，并设置finish_reason为其他值
+        output = {"finish_reason": "other_reason"}
+        # 调用calc_finish_reason方法
+        result = serving_completion.calc_finish_reason(None, 100, output)
+        # 断言结果为"stop"
+        assert result == "stop"
+
+    def test_calc_finish_reason_length(self):
+        # 创建一个模拟的engine_client
+        engine_client = Mock()
+        # 创建一个OpenAIServingCompletion实例
+        serving_completion = OpenAIServingCompletion(engine_client, "pid", "ips")
+        # 创建一个模拟的output
+        output = {}
+        # 调用calc_finish_reason方法
+        result = serving_completion.calc_finish_reason(100, 100, output)
+        # 断言结果为"length"
+        assert result == "length"
+
+
+if __name__ == "__main__":
+    unittest.main()