[2/N] API Server: Avoid ulimit footgun (#11530)

vllm/entrypoints/api_server.py

@@ -21,7 +21,7 @@ from vllm.entrypoints.utils import with_cancellation
 from vllm.logger import init_logger
 from vllm.sampling_params import SamplingParams
 from vllm.usage.usage_lib import UsageContext
-from vllm.utils import FlexibleArgumentParser, random_uuid
+from vllm.utils import FlexibleArgumentParser, random_uuid, set_ulimit
 from vllm.version import __version__ as VLLM_VERSION
 
 logger = init_logger("vllm.entrypoints.api_server")
@@ -119,6 +119,8 @@ async def run_server(args: Namespace,
     logger.info("vLLM API server version %s", VLLM_VERSION)
     logger.info("args: %s", args)
 
+    set_ulimit()
+
     app = await init_app(args, llm_engine)
     assert engine is not None
 

vllm/entrypoints/openai/api_server.py

@@ -68,7 +68,7 @@ from vllm.entrypoints.utils import with_cancellation
 from vllm.logger import init_logger
 from vllm.usage.usage_lib import UsageContext
 from vllm.utils import (FlexibleArgumentParser, get_open_zmq_ipc_path,
-                        is_valid_ipv6_address)
+                        is_valid_ipv6_address, set_ulimit)
 from vllm.version import __version__ as VLLM_VERSION
 
 TIMEOUT_KEEP_ALIVE = 5  # seconds
@@ -727,6 +727,10 @@ async def run_server(args, **uvicorn_kwargs) -> None:
     sock_addr = (args.host or "", args.port)
     sock = create_server_socket(sock_addr)
 
+    # workaround to avoid footguns where uvicorn drops requests with too
+    # many concurrent requests active
+    set_ulimit()
+
     def signal_handler(*_) -> None:
         # Interrupt server on sigterm while initializing
         raise KeyboardInterrupt("terminated")