Fix ignore_eos parameter when loading a chat template (#5264)

python/sglang/srt/openai_api/adapter.py

@@ -995,7 +995,8 @@ def v1_chat_generate_request(
                 image_data = conv.image_data
                 audio_data = conv.audio_data
                 modalities = conv.modalities
-                stop = conv.stop_str or []
+                stop = conv.stop_str or [] if not request.ignore_eos else []
+
                 if request.stop:
                     if isinstance(request.stop, str):
                         stop.append(request.stop)

test/srt/test_openai_server.py

@@ -676,5 +676,79 @@ class TestOpenAIEmbedding(CustomTestCase):
         self.assertTrue(len(response.data[1].embedding) > 0)
 
 
+class TestOpenAIServerIgnoreEOS(CustomTestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = DEFAULT_SMALL_MODEL_NAME_FOR_TEST
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.api_key = "sk-123456"
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            api_key=cls.api_key,
+            other_args=["--chat-template=llama_3_vision"],
+        )
+        cls.base_url += "/v1"
+        cls.tokenizer = get_tokenizer(DEFAULT_SMALL_MODEL_NAME_FOR_TEST)
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_ignore_eos(self):
+        """
+        Test that ignore_eos=True allows generation to continue beyond EOS token
+        and reach the max_tokens limit.
+        """
+        client = openai.Client(api_key=self.api_key, base_url=self.base_url)
+
+        max_tokens = 200
+
+        response_default = client.chat.completions.create(
+            model=self.model,
+            messages=[
+                {"role": "system", "content": "You are a helpful assistant."},
+                {"role": "user", "content": "Count from 1 to 20."},
+            ],
+            temperature=0,
+            max_tokens=max_tokens,
+            extra_body={"ignore_eos": False},
+        )
+
+        response_ignore_eos = client.chat.completions.create(
+            model=self.model,
+            messages=[
+                {"role": "system", "content": "You are a helpful assistant."},
+                {"role": "user", "content": "Count from 1 to 20."},
+            ],
+            temperature=0,
+            max_tokens=max_tokens,
+            extra_body={"ignore_eos": True},
+        )
+
+        default_tokens = len(
+            self.tokenizer.encode(response_default.choices[0].message.content)
+        )
+        ignore_eos_tokens = len(
+            self.tokenizer.encode(response_ignore_eos.choices[0].message.content)
+        )
+
+        # Check if ignore_eos resulted in more tokens or exactly max_tokens
+        # The ignore_eos response should either:
+        # 1. Have more tokens than the default response (if default stopped at EOS before max_tokens)
+        # 2. Have exactly max_tokens (if it reached the max_tokens limit)
+        self.assertTrue(
+            ignore_eos_tokens > default_tokens or ignore_eos_tokens >= max_tokens,
+            f"ignore_eos did not generate more tokens: {ignore_eos_tokens} vs {default_tokens}",
+        )
+
+        self.assertEqual(
+            response_ignore_eos.choices[0].finish_reason,
+            "length",
+            f"Expected finish_reason='length' for ignore_eos=True, got {response_ignore_eos.choices[0].finish_reason}",
+        )
+
+
 if __name__ == "__main__":
     unittest.main()