[Doc][CI/Build] Update docs and tests to use `vllm serve` (#6431)

tests/entrypoints/openai/test_vision.py

@@ -23,17 +23,17 @@ TEST_IMAGE_URLS = [
 
 @pytest.fixture(scope="module")
 def server():
-    with RemoteOpenAIServer([
-            "--model",
-            MODEL_NAME,
-            "--dtype",
-            "bfloat16",
-            "--max-model-len",
-            "4096",
-            "--enforce-eager",
-            "--chat-template",
-            str(LLAVA_CHAT_TEMPLATE),
-    ]) as remote_server:
+    args = [
+        "--dtype",
+        "bfloat16",
+        "--max-model-len",
+        "4096",
+        "--enforce-eager",
+        "--chat-template",
+        str(LLAVA_CHAT_TEMPLATE),
+    ]
+
+    with RemoteOpenAIServer(MODEL_NAME, args) as remote_server:
         yield remote_server
 
 

tests/tensorizer_loader/test_tensorizer.py

@@ -214,12 +214,12 @@ def test_openai_apiserver_with_tensorizer(vllm_runner, tmp_path):
 
     ## Start OpenAI API server
     openai_args = [
-        "--model", model_ref, "--dtype", "float16", "--load-format",
+        "--dtype", "float16", "--load-format",
         "tensorizer", "--model-loader-extra-config",
         json.dumps(model_loader_extra_config),
     ]
 
-    with RemoteOpenAIServer(openai_args) as server:
+    with RemoteOpenAIServer(model_ref, openai_args) as server:
         print("Server ready.")
 
         client = server.get_client()

tests/utils.py

@@ -49,7 +49,13 @@ class RemoteOpenAIServer:
     DUMMY_API_KEY = "token-abc123"  # vLLM's OpenAI server does not need API key
     MAX_SERVER_START_WAIT_S = 600  # wait for server to start for 60 seconds
 
-    def __init__(self, cli_args: List[str], *, auto_port: bool = True) -> None:
+    def __init__(
+        self,
+        model: str,
+        cli_args: List[str],
+        *,
+        auto_port: bool = True,
+    ) -> None:
         if auto_port:
             if "-p" in cli_args or "--port" in cli_args:
                 raise ValueError("You have manually specified the port"
@@ -68,12 +74,10 @@ class RemoteOpenAIServer:
         # the current process might initialize cuda,
         # to be safe, we should use spawn method
         env['VLLM_WORKER_MULTIPROC_METHOD'] = 'spawn'
-        self.proc = subprocess.Popen(
-            [sys.executable, "-m", "vllm.entrypoints.openai.api_server"] +
-            cli_args,
-            env=env,
-            stdout=sys.stdout,
-            stderr=sys.stderr)
+        self.proc = subprocess.Popen(["vllm", "serve"] + [model] + cli_args,
+                                     env=env,
+                                     stdout=sys.stdout,
+                                     stderr=sys.stderr)
         self._wait_for_server(url=self.url_for("health"),
                               timeout=self.MAX_SERVER_START_WAIT_S)