[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")

vllm/utils.py

@@ -12,6 +12,7 @@ import inspect
 import ipaddress
 import os
 import re
+import resource
 import signal
 import socket
 import subprocess
@@ -1818,3 +1819,20 @@ def memory_profiling(
     result.non_torch_increase_in_bytes = current_cuda_memory_bytes - baseline_memory_in_bytes - weights_memory_in_bytes - diff.torch_memory_in_bytes  # noqa
     result.profile_time = diff.timestamp
     result.non_kv_cache_memory_in_bytes = result.non_torch_increase_in_bytes + result.torch_peak_increase_in_bytes + result.weights_memory_in_bytes  # noqa
+
+
+# Adapted from: https://github.com/sgl-project/sglang/blob/f46f394f4d4dbe4aae85403dec006199b34d2840/python/sglang/srt/utils.py#L630 # noqa: E501Curre
+def set_ulimit(target_soft_limit=65535):
+    resource_type = resource.RLIMIT_NOFILE
+    current_soft, current_hard = resource.getrlimit(resource_type)
+
+    if current_soft < target_soft_limit:
+        try:
+            resource.setrlimit(resource_type,
+                               (target_soft_limit, current_hard))
+        except ValueError as e:
+            logger.warning(
+                "Found ulimit of %s and failed to automatically increase"
+                "with error %s. This can cause fd limit errors like"
+                "`OSError: [Errno 24] Too many open files`. Consider "
+                "increasing with ulimit -n", current_soft, e)