【Bugfix】修复2.1分支上0.3B模型性能大幅下降 (#3624)

* 恢复异步方法。
【BugFix】completion接口echo回显支持 (#3245)

* wenxin-tools-511,修复v1/completion无法回显的问题。

* 支持多prompt的回显

* 支持多prompt情况下的流式回显

* 补充了 completion 接口支持 echo 的单元测试

* pre-commit

* 移除了多余的test文件

* 修复了completion接口echo支持的单测方法

* 补充了单元测试文件

* 补充单测

* unittest

* 补充单测

* 修复单测

* 删除不必要的assert.

* 重新提交

* 更新测试方法

* ut

* 验证是否是正确思路单测

* 验证是否是正确思路单测

* 验证是否是正确思路单测3

* 优化单测代码，有针对性地缩小单测范围。

* 优化单测代码2，有针对性地缩小单测范围。

* 优化单测代码3，有针对性地缩小单测范围。

* support 'echo' in chat/completion.

* update

* update

* update

* update

* update

* update

* 补充了关于tokenid的单元测试

* update

* 修正index错误

* 修正index错误

* [Bugfix] Significant performance degradation of 0.3B model on branch 2.1

fastdeploy/entrypoints/openai/serving_completion.py

@@ -240,7 +240,7 @@ class OpenAIServingCompletion:
                 dealer.close()
                 self.engine_client.semaphore.release()
 
-    def _echo_back_prompt(self, request, res, idx):
+    async def _echo_back_prompt(self, request, res, idx):
         if res["outputs"].get("send_idx", -1) == 0 and request.echo:
             if isinstance(request.prompt, list):
                 prompt_text = request.prompt[idx]
@@ -346,7 +346,7 @@ class OpenAIServingCompletion:
                     else:
                         arrival_time = res["metrics"]["arrival_time"] - inference_start_time[idx]
 
-                    self._echo_back_prompt(request, res, idx)
+                    await self._echo_back_prompt(request, res, idx)
                     output = res["outputs"]
                     output_top_logprobs = output["top_logprobs"]
                     logprobs_res: Optional[CompletionLogprobs] = None
@@ -471,7 +471,6 @@ class OpenAIServingCompletion:
             else:
                 token_ids = output["token_ids"]
                 output_text = output["text"]
-
             finish_reason = self.calc_finish_reason(request.max_tokens, final_res["output_token_ids"], output, False)
 
             choice_data = CompletionResponseChoice(

test/entrypoints/openai/test_completion_echo.py

@@ -7,13 +7,22 @@ from fastdeploy.entrypoints.openai.serving_completion import (
 )
 
 
-class TestCompletionEcho(unittest.TestCase):
+class YourClass:
+    async def _1(self, a, b, c):
+        if b["outputs"].get("send_idx", -1) == 0 and a.echo:
+            if isinstance(a.prompt, list):
+                text = a.prompt[c]
+            else:
+                text = a.prompt
+            b["outputs"]["text"] = text + (b["outputs"]["text"] or "")
+
+
+class TestCompletionEcho(unittest.IsolatedAsyncioTestCase):
     def setUp(self):
         self.mock_engine = MagicMock()
         self.completion_handler = None
 
     def test_single_prompt_non_streaming(self):
-        """测试单prompt非流式响应"""
         self.completion_handler = OpenAIServingCompletion(self.mock_engine, pid=123, ips=None, max_waiting_time=30)
 
         request = CompletionRequest(prompt="test prompt", max_tokens=10, echo=True, logprobs=1)
@@ -42,8 +51,7 @@ class TestCompletionEcho(unittest.TestCase):
 
         self.assertEqual(response.choices[0].text, "test prompt generated text")
 
-    def test_echo_back_prompt_and_streaming(self):
+    async def test_echo_back_prompt_and_streaming(self):
-        """测试_echo_back_prompt方法和流式响应的prompt拼接逻辑"""
         self.completion_handler = OpenAIServingCompletion(self.mock_engine, pid=123, ips=None, max_waiting_time=30)
 
         request = CompletionRequest(prompt="test prompt", max_tokens=10, stream=True, echo=True)
@@ -57,7 +65,7 @@ class TestCompletionEcho(unittest.TestCase):
 
             mock_echo.side_effect = mock_echo_side_effect
 
-            self.completion_handler._echo_back_prompt(request, mock_response, 0)
+            await self.completion_handler._echo_back_prompt(request, mock_response, 0)
 
             mock_echo.assert_called_once_with(request, mock_response, 0)
 
@@ -65,7 +73,6 @@ class TestCompletionEcho(unittest.TestCase):
             self.assertEqual(request.prompt, "test prompt")
 
     def test_multi_prompt_non_streaming(self):
-        """测试多prompt非流式响应"""
         self.completion_handler = OpenAIServingCompletion(self.mock_engine, pid=123, ips=None, max_waiting_time=30)
 
         request = CompletionRequest(prompt=["prompt1", "prompt2"], max_tokens=10, echo=True)
@@ -97,7 +104,7 @@ class TestCompletionEcho(unittest.TestCase):
         self.assertEqual(response.choices[0].text, "prompt1 response1")
         self.assertEqual(response.choices[1].text, "prompt2 response2")
 
-    def test_multi_prompt_streaming(self):
+    async def test_multi_prompt_streaming(self):
         self.completion_handler = OpenAIServingCompletion(self.mock_engine, pid=123, ips=None, max_waiting_time=30)
 
         request = CompletionRequest(prompt=["prompt1", "prompt2"], max_tokens=10, stream=True, echo=True)
@@ -114,8 +121,8 @@ class TestCompletionEcho(unittest.TestCase):
 
             mock_echo.side_effect = mock_echo_side_effect
 
-            self.completion_handler._echo_back_prompt(request, mock_responses[0], 0)
+            await self.completion_handler._echo_back_prompt(request, mock_responses[0], 0)
-            self.completion_handler._echo_back_prompt(request, mock_responses[1], 1)
+            await self.completion_handler._echo_back_prompt(request, mock_responses[1], 1)
 
             self.assertEqual(mock_echo.call_count, 2)
             mock_echo.assert_any_call(request, mock_responses[0], 0)
@@ -125,41 +132,40 @@ class TestCompletionEcho(unittest.TestCase):
             self.assertEqual(mock_responses[1]["outputs"]["text"], "prompt2 response2")
             self.assertEqual(request.prompt, ["prompt1", "prompt2"])
 
-    def test_echo_back_prompt_and_streaming1(self):
+    async def test_echo_back_prompt_and_streaming1(self):
         request = CompletionRequest(echo=True, prompt=["Hello", "World"])
         res = {"outputs": {"send_idx": 0, "text": "!"}}
         idx = 0
 
         instance = OpenAIServingCompletion(self.mock_engine, pid=123, ips=None, max_waiting_time=30)
-        instance._echo_back_prompt(request, res, idx)
+        await instance._echo_back_prompt(request, res, idx)
         self.assertEqual(res["outputs"]["text"], "Hello!")
 
-    def test_1_prompt_is_string_and_send_idx_is_0(self):
+    async def test_1_prompt_is_string_and_send_idx_is_0(self):
         request = CompletionRequest(echo=True, prompt="Hello")
         res = {"outputs": {"send_idx": 0, "text": "!"}}
         idx = 0
 
         instance = OpenAIServingCompletion(self.mock_engine, pid=123, ips=None, max_waiting_time=30)
-        instance._echo_back_prompt(request, res, idx)
+        await instance._echo_back_prompt(request, res, idx)
         self.assertEqual(res["outputs"]["text"], "Hello!")
 
-    def test_1_send_idx_is_not_0(self):
+    async def test_1_send_idx_is_not_0(self):
         request = CompletionRequest(echo=True, prompt="Hello")
         res = {"outputs": {"send_idx": 1, "text": "!"}}
         idx = 0
 
         instance = OpenAIServingCompletion(self.mock_engine, pid=123, ips=None, max_waiting_time=30)
-        instance._echo_back_prompt(request, res, idx)
+        await instance._echo_back_prompt(request, res, idx)
         self.assertEqual(res["outputs"]["text"], "!")
 
-    def test_1_echo_is_false(self):
+    async def test_1_echo_is_false(self):
-        """测试echo为False时，_echo_back_prompt不拼接prompt"""
         request = CompletionRequest(echo=False, prompt="Hello")
         res = {"outputs": {"send_idx": 0, "text": "!"}}
         idx = 0
 
         instance = OpenAIServingCompletion(self.mock_engine, pid=123, ips=None, max_waiting_time=30)
-        instance._echo_back_prompt(request, res, idx)
+        await instance._echo_back_prompt(request, res, idx)
         self.assertEqual(res["outputs"]["text"], "!")