fix: correct stream response when enable_thinking is set to false (#5881)

minleminzui · web-flow · commit 256c4c25194c · 2025-04-30T19:44:37.000-07:00
diff --git a/python/sglang/srt/openai_api/adapter.py b/python/sglang/srt/openai_api/adapter.py
@@ -899,6 +899,24 @@ async def generate_stream_resp():
     return response
 
 
+def _get_enable_thinking_from_request(request_obj):
+    """Extracts the 'enable_thinking' flag from request chat_template_kwargs.
+
+    Args:
+        request_obj: The request object (or an item from a list of requests).
+
+    Returns:
+        The boolean value of 'enable_thinking' if found and not True, otherwise True.
+    """
+    if (
+        hasattr(request_obj, "chat_template_kwargs")
+        and request_obj.chat_template_kwargs
+        and request_obj.chat_template_kwargs.get("enable_thinking") is not None
+    ):
+        return request_obj.chat_template_kwargs.get("enable_thinking")
+    return True
+
+
 def v1_chat_generate_request(
     all_requests: List[ChatCompletionRequest],
     tokenizer_manager,
@@ -1263,31 +1281,16 @@ def v1_chat_generate_response(
         tool_calls = None
         text = ret_item["text"]
 
-        enable_thinking = True
         if isinstance(request, list):
             tool_choice = request[idx].tool_choice
             tools = request[idx].tools
             separate_reasoning = request[idx].separate_reasoning
-
-            if (
-                request[idx].chat_template_kwargs
-                and request[idx].chat_template_kwargs.get("enable_thinking") is not None
-            ):
-                enable_thinking = request[idx].chat_template_kwargs.get(
-                    "enable_thinking", True
-                )
+            enable_thinking = _get_enable_thinking_from_request(request[idx])
         else:
             tool_choice = request.tool_choice
             tools = request.tools
             separate_reasoning = request.separate_reasoning
-
-            if (
-                request.chat_template_kwargs
-                and request.chat_template_kwargs.get("enable_thinking") is not None
-            ):
-                enable_thinking = request.chat_template_kwargs.get(
-                    "enable_thinking", True
-                )
+            enable_thinking = _get_enable_thinking_from_request(request)
 
         reasoning_text = None
         if reasoning_parser and separate_reasoning and enable_thinking:
@@ -1526,9 +1529,12 @@ async def generate_stream_resp():
                     delta = text[len(stream_buffer) :]
                     new_stream_buffer = stream_buffer + delta
 
+                    enable_thinking = _get_enable_thinking_from_request(request)
+
                     if (
                         tokenizer_manager.server_args.reasoning_parser
                         and request.separate_reasoning
+                        and enable_thinking
                     ):
                         if index not in reasoning_parser_dict:
                             reasoning_parser_dict[index] = ReasoningParser(
diff --git a/python/sglang/test/test_utils.py b/python/sglang/test/test_utils.py
@@ -69,6 +69,7 @@
 DEFAULT_AWQ_MOE_MODEL_NAME_FOR_TEST = (
     "hugging-quants/Mixtral-8x7B-Instruct-v0.1-AWQ-INT4"
 )
+DEFAULT_ENABLE_THINKING_MODEL_NAME_FOR_TEST = "Qwen/Qwen3-30B-A3B"
 
 # Nightly tests
 DEFAULT_MODEL_NAME_FOR_NIGHTLY_EVAL_TP1 = "meta-llama/Llama-3.1-8B-Instruct,mistralai/Mistral-7B-Instruct-v0.3,deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct,google/gemma-2-27b-it"
diff --git a/test/srt/run_suite.py b/test/srt/run_suite.py
@@ -59,6 +59,7 @@ class TestFile:
         TestFile("test_pytorch_sampling_backend.py", 66),
         TestFile("test_radix_attention.py", 167),
         TestFile("test_reasoning_content.py", 89),
+        TestFile("test_enable_thinking.py", 70),
         TestFile("test_regex_constrained.py", 64),
         TestFile("test_release_memory_occupation.py", 44),
         TestFile("test_request_length_validation.py", 31),
diff --git a/test/srt/test_enable_thinking.py b/test/srt/test_enable_thinking.py
@@ -0,0 +1,186 @@
+"""
+Usage:
+python3 -m unittest test_enable_thinking.TestEnableThinking.test_chat_completion_with_reasoning
+python3 -m unittest test_enable_thinking.TestEnableThinking.test_chat_completion_without_reasoning
+python3 -m unittest test_enable_thinking.TestEnableThinking.test_stream_chat_completion_with_reasoning
+python3 -m unittest test_enable_thinking.TestEnableThinking.test_stream_chat_completion_without_reasoning
+"""
+
+import asyncio
+import json
+import os
+import sys
+import time
+import unittest
+
+import requests
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.test_utils import (
+    DEFAULT_ENABLE_THINKING_MODEL_NAME_FOR_TEST,
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    popen_launch_server,
+)
+
+
+class TestEnableThinking(CustomTestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = DEFAULT_ENABLE_THINKING_MODEL_NAME_FOR_TEST
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.api_key = "sk-1234"
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            api_key=cls.api_key,
+            other_args=[
+                "--reasoning-parser",
+                "qwen3",
+            ],
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_chat_completion_with_reasoning(self):
+        # Test non-streaming with "enable_thinking": True, reasoning_content should not be empty
+        client = requests.post(
+            f"{self.base_url}/v1/chat/completions",
+            headers={"Authorization": f"Bearer {self.api_key}"},
+            json={
+                "model": self.model,
+                "messages": [{"role": "user", "content": "Hello"}],
+                "temperature": 0,
+                "separate_reasoning": True,
+                "chat_template_kwargs": {"enable_thinking": True},
+            },
+        )
+
+        self.assertEqual(client.status_code, 200, f"Failed with: {client.text}")
+        data = client.json()
+
+        self.assertIn("choices", data)
+        self.assertTrue(len(data["choices"]) > 0)
+        self.assertIn("message", data["choices"][0])
+        self.assertIn("reasoning_content", data["choices"][0]["message"])
+        self.assertIsNotNone(data["choices"][0]["message"]["reasoning_content"])
+
+    def test_chat_completion_without_reasoning(self):
+        # Test non-streaming with "enable_thinking": False, reasoning_content should be empty
+        client = requests.post(
+            f"{self.base_url}/v1/chat/completions",
+            headers={"Authorization": f"Bearer {self.api_key}"},
+            json={
+                "model": self.model,
+                "messages": [{"role": "user", "content": "Hello"}],
+                "temperature": 0,
+                "separate_reasoning": True,
+                "chat_template_kwargs": {"enable_thinking": False},
+            },
+        )
+
+        self.assertEqual(client.status_code, 200, f"Failed with: {client.text}")
+        data = client.json()
+
+        self.assertIn("choices", data)
+        self.assertTrue(len(data["choices"]) > 0)
+        self.assertIn("message", data["choices"][0])
+
+        if "reasoning_content" in data["choices"][0]["message"]:
+            self.assertIsNone(data["choices"][0]["message"]["reasoning_content"])
+
+    def test_stream_chat_completion_with_reasoning(self):
+        # Test streaming with "enable_thinking": True, reasoning_content should not be empty
+        response = requests.post(
+            f"{self.base_url}/v1/chat/completions",
+            headers={"Authorization": f"Bearer {self.api_key}"},
+            json={
+                "model": self.model,
+                "messages": [{"role": "user", "content": "Hello"}],
+                "temperature": 0,
+                "separate_reasoning": True,
+                "stream": True,
+                "chat_template_kwargs": {"enable_thinking": True},
+            },
+            stream=True,
+        )
+
+        self.assertEqual(response.status_code, 200, f"Failed with: {response.text}")
+
+        has_reasoning = False
+        has_content = False
+
+        print("\n=== Stream With Reasoning ===")
+        for line in response.iter_lines():
+            if line:
+                line = line.decode("utf-8")
+                if line.startswith("data:") and not line.startswith("data: [DONE]"):
+                    data = json.loads(line[6:])
+                    if "choices" in data and len(data["choices"]) > 0:
+                        delta = data["choices"][0].get("delta", {})
+
+                        if "reasoning_content" in delta and delta["reasoning_content"]:
+                            has_reasoning = True
+
+                        if "content" in delta and delta["content"]:
+                            has_content = True
+
+        self.assertTrue(
+            has_reasoning,
+            "The reasoning content is not included in the stream response",
+        )
+        self.assertTrue(
+            has_content, "The stream response does not contain normal content"
+        )
+
+    def test_stream_chat_completion_without_reasoning(self):
+        # Test streaming with "enable_thinking": False, reasoning_content should  be empty
+        response = requests.post(
+            f"{self.base_url}/v1/chat/completions",
+            headers={"Authorization": f"Bearer {self.api_key}"},
+            json={
+                "model": self.model,
+                "messages": [{"role": "user", "content": "Hello"}],
+                "temperature": 0,
+                "separate_reasoning": True,
+                "stream": True,
+                "chat_template_kwargs": {"enable_thinking": False},
+            },
+            stream=True,
+        )
+
+        self.assertEqual(response.status_code, 200, f"Failed with: {response.text}")
+
+        has_reasoning = False
+        has_content = False
+
+        print("\n=== Stream Without Reasoning ===")
+        for line in response.iter_lines():
+            if line:
+                line = line.decode("utf-8")
+                if line.startswith("data:") and not line.startswith("data: [DONE]"):
+                    data = json.loads(line[6:])
+                    if "choices" in data and len(data["choices"]) > 0:
+                        delta = data["choices"][0].get("delta", {})
+
+                        if "reasoning_content" in delta and delta["reasoning_content"]:
+                            has_reasoning = True
+
+                        if "content" in delta and delta["content"]:
+                            has_content = True
+
+        self.assertFalse(
+            has_reasoning,
+            "The reasoning content should not be included in the stream response",
+        )
+        self.assertTrue(
+            has_content, "The stream response does not contain normal content"
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()

Original file line number	Diff line number	Diff line change
`@@ -69,6 +69,7 @@`
`69`	`69`	`DEFAULT_AWQ_MOE_MODEL_NAME_FOR_TEST = (`
`70`	`70`	`"hugging-quants/Mixtral-8x7B-Instruct-v0.1-AWQ-INT4"`
`71`	`71`	`)`
	`72`	`+DEFAULT_ENABLE_THINKING_MODEL_NAME_FOR_TEST = "Qwen/Qwen3-30B-A3B"`
`72`	`73`
`73`	`74`	`# Nightly tests`
`74`	`75`	`DEFAULT_MODEL_NAME_FOR_NIGHTLY_EVAL_TP1 = "meta-llama/Llama-3.1-8B-Instruct,mistralai/Mistral-7B-Instruct-v0.3,deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct,google/gemma-2-27b-it"`