[BugFix] Fix frontend multiprocessing hang (vllm-project#7217)

Signed-off-by: Max de Bayser <[email protected]>
Co-authored-by: Robert Shaw <[email protected]>
Signed-off-by: Alvant <[email protected]>

Author: 2 people

tests/entrypoints/openai/test_mp_crash.py

@@ -0,0 +1,35 @@
+from typing import Any
+
+import pytest
+
+from vllm.engine.async_llm_engine import AsyncLLMEngine
+from vllm.entrypoints.openai.api_server import build_async_engine_client
+from vllm.entrypoints.openai.cli_args import make_arg_parser
+from vllm.utils import FlexibleArgumentParser
+
+
+def crashing_from_engine_args(
+    cls,
+    engine_args: Any = None,
+    start_engine_loop: Any = None,
+    usage_context: Any = None,
+    stat_loggers: Any = None,
+) -> "AsyncLLMEngine":
+    raise Exception("foo")
+
+
+@pytest.mark.asyncio
+async def test_mp_crash_detection(monkeypatch):
+
+    with pytest.raises(RuntimeError) as excinfo, monkeypatch.context() as m:
+        m.setattr(AsyncLLMEngine, "from_engine_args",
+                  crashing_from_engine_args)
+        parser = FlexibleArgumentParser(
+            description="vLLM's remote OpenAI server.")
+        parser = make_arg_parser(parser)
+        args = parser.parse_args([])
+
+        async with build_async_engine_client(args):
+            pass
+    assert "The server process died before responding to the readiness probe"\
+          in str(excinfo.value)

vllm/entrypoints/openai/api_server.py

@@ -120,9 +120,18 @@ async def build_async_engine_client(args) -> AsyncIterator[AsyncEngineClient]:
 
         # Build RPCClient, which conforms to AsyncEngineClient Protocol.
         async_engine_client = AsyncEngineRPCClient(rpc_path)
-        await async_engine_client.setup()
 
         try:
+            while True:
+                try:
+                    await async_engine_client.setup()
+                    break
+                except TimeoutError as e:
+                    if not rpc_server_process.is_alive():
+                        raise RuntimeError(
+                            "The server process died before "
+                            "responding to the readiness probe") from e
+
             yield async_engine_client
         finally:
             # Ensure rpc server process was terminated

vllm/entrypoints/openai/rpc/client.py

@@ -18,6 +18,9 @@
 from vllm.sampling_params import SamplingParams
 from vllm.transformers_utils.tokenizer_group import init_tokenizer_from_configs
 
+# Time to wait before checking it the server process is alive.
+SERVER_START_TIMEOUT_MS = 1000
+
 
 class AsyncEngineRPCClient:
 
@@ -61,7 +64,16 @@ def socket(self):
             socket.connect(self.rpc_path)
             yield socket
         finally:
-            socket.close()
+            # linger == 0 means discard unsent messages
+            # when the socket is closed. This is necessary
+            # because otherwise self.context.destroy() will
+            # wait for 30 seconds until unsent messages are
+            # received, which is impossible if the server
+            # crashed. In the absence of a server crash we
+            # always expect a response before closing the
+            # socket anyway.
+            # Reference: http://api.zeromq.org/4-2:zmq-setsockopt#toc24
+            socket.close(linger=0)
 
     async def _send_get_data_rpc_request(self, request: RPCUtilityRequest,
                                          expected_type: Any,
@@ -85,14 +97,19 @@ async def _send_get_data_rpc_request(self, request: RPCUtilityRequest,
 
         return data
 
-    async def _send_one_way_rpc_request(self, request: RPC_REQUEST_TYPE,
-                                        error_message: str):
+    async def _send_one_way_rpc_request(self,
+                                        request: RPC_REQUEST_TYPE,
+                                        error_message: str,
+                                        timeout: Optional[int] = None):
         """Send one-way RPC request to trigger an action."""
         with self.socket() as socket:
             # Ping RPC Server with request.
             await socket.send(cloudpickle.dumps(request))
 
             # Await acknowledgement from RPCServer.
+            if timeout is not None and await socket.poll(timeout=timeout) == 0:
+                raise TimeoutError(f"server didn't reply within {timeout} ms")
+
             response = cloudpickle.loads(await socket.recv())
 
         if not isinstance(response, str) or response != VLLM_RPC_SUCCESS_STR:
@@ -117,7 +134,8 @@ async def wait_for_server(self):
 
         await self._send_one_way_rpc_request(
             request=RPCUtilityRequest.IS_SERVER_READY,
-            error_message="Unable to start RPC Server.")
+            error_message="Unable to start RPC Server.",
+            timeout=SERVER_START_TIMEOUT_MS)
 
     async def _get_model_config_rpc(self) -> ModelConfig:
         """Get the ModelConfig object from the RPC Server"""