[Bugfix] Fix embedding model hangs with --enable-metrics (#2822)

test/srt/test_openai_server.py

@@ -14,6 +14,7 @@ import openai
 from sglang.srt.hf_transformers_utils import get_tokenizer
 from sglang.srt.utils import kill_process_tree
 from sglang.test.test_utils import (
+    DEFAULT_SMALL_EMBEDDING_MODEL_NAME_FOR_TEST,
     DEFAULT_SMALL_MODEL_NAME_FOR_TEST,
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
     DEFAULT_URL_FOR_TEST,
@@ -675,5 +676,45 @@ class TestOpenAIServerEBNF(unittest.TestCase):
         ), "Function name should be add for the above response"
 
 
+class TestOpenAIEmbedding(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = DEFAULT_SMALL_EMBEDDING_MODEL_NAME_FOR_TEST
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.api_key = "sk-123456"
+
+        # Configure embedding-specific args
+        other_args = ["--is-embedding", "--enable-metrics"]
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            api_key=cls.api_key,
+            other_args=other_args,
+        )
+        cls.base_url += "/v1"
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_embedding_single(self):
+        """Test single embedding request"""
+        client = openai.Client(api_key=self.api_key, base_url=self.base_url)
+        response = client.embeddings.create(model=self.model, input="Hello world")
+        self.assertEqual(len(response.data), 1)
+        self.assertTrue(len(response.data[0].embedding) > 0)
+
+    def test_embedding_batch(self):
+        """Test batch embedding request"""
+        client = openai.Client(api_key=self.api_key, base_url=self.base_url)
+        response = client.embeddings.create(
+            model=self.model, input=["Hello world", "Test text"]
+        )
+        self.assertEqual(len(response.data), 2)
+        self.assertTrue(len(response.data[0].embedding) > 0)
+        self.assertTrue(len(response.data[1].embedding) > 0)
+
+
 if __name__ == "__main__":
     unittest.main()