[Core] V1: Use multiprocessing by default (#11074)

Signed-off-by: Russell Bryant <rbryant@redhat.com>
This commit is contained in:
Russell Bryant
2024-12-13 16:27:32 -08:00
committed by GitHub
parent 0d8451c3a4
commit 4863e5fba5
10 changed files with 299 additions and 17 deletions
+4
View File
@@ -232,6 +232,10 @@ class LLM:
self.request_counter = Counter()
def __del__(self):
if self.llm_engine and hasattr(self.llm_engine, "shutdown"):
self.llm_engine.shutdown()
@staticmethod
def get_engine_class() -> Type[LLMEngine]:
if envs.VLLM_USE_V1:
+2 -2
View File
@@ -69,7 +69,7 @@ if TYPE_CHECKING:
VLLM_SKIP_P2P_CHECK: bool = False
VLLM_DISABLED_KERNELS: List[str] = []
VLLM_USE_V1: bool = False
VLLM_ENABLE_V1_MULTIPROCESSING: bool = False
VLLM_ENABLE_V1_MULTIPROCESSING: bool = True
VLLM_LOG_BATCHSIZE_INTERVAL: float = -1
@@ -460,7 +460,7 @@ environment_variables: Dict[str, Callable[[], Any]] = {
# If set, enable multiprocessing in LLM for the V1 code path.
"VLLM_ENABLE_V1_MULTIPROCESSING":
lambda: bool(int(os.getenv("VLLM_ENABLE_V1_MULTIPROCESSING", "0"))),
lambda: bool(int(os.getenv("VLLM_ENABLE_V1_MULTIPROCESSING", "1"))),
"VLLM_LOG_BATCHSIZE_INTERVAL":
lambda: float(os.getenv("VLLM_LOG_BATCHSIZE_INTERVAL", "-1")),
}
+14 -6
View File
@@ -274,7 +274,20 @@ def _add_prefix(file: TextIO, worker_name: str, pid: int) -> None:
file.write = write_with_prefix # type: ignore[method-assign]
def _check_multiproc_method():
if (cuda_is_initialized()
and os.environ.get("VLLM_WORKER_MULTIPROC_METHOD") != "spawn"):
logger.warning("CUDA was previously initialized. We must use "
"the `spawn` multiprocessing start method. Setting "
"VLLM_WORKER_MULTIPROC_METHOD to 'spawn'. "
"See https://docs.vllm.ai/en/latest/getting_started/"
"debugging.html#python-multiprocessing "
"for more information.")
os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
def get_mp_context():
_check_multiproc_method()
mp_method = envs.VLLM_WORKER_MULTIPROC_METHOD
return multiprocessing.get_context(mp_method)
@@ -284,12 +297,7 @@ def set_multiprocessing_worker_envs(parallel_config):
in a multiprocessing environment. This should be called by the parent
process before worker processes are created"""
if (cuda_is_initialized()
and os.environ.get("VLLM_WORKER_MULTIPROC_METHOD") != "spawn"):
logger.warning("CUDA was previously initialized. We must use "
"the `spawn` multiprocessing start method. Setting "
"VLLM_WORKER_MULTIPROC_METHOD to 'spawn'.")
os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
_check_multiproc_method()
# Configure thread parallelism if OMP_NUM_THREADS isn't set
#
+2 -6
View File
@@ -1,4 +1,3 @@
import multiprocessing
import pickle
import queue
import signal
@@ -13,6 +12,7 @@ import zmq.asyncio
from msgspec import msgpack
from vllm.config import CacheConfig, VllmConfig
from vllm.executor.multiproc_worker_utils import get_mp_context
from vllm.logger import init_logger
from vllm.usage.usage_lib import UsageContext
from vllm.v1.core.scheduler import Scheduler
@@ -210,11 +210,7 @@ class EngineCoreProc(EngineCore):
output_path: str,
ready_path: str,
) -> EngineCoreProcHandle:
# The current process might have CUDA context,
# so we need to spawn a new process.
# NOTE(rob): this is a problem for using EngineCoreProc w/
# LLM, since we need a if __name__ == "__main__" guard.
context = multiprocessing.get_context("spawn")
context = get_mp_context()
process_kwargs = {
"input_path": input_path,
+9 -2
View File
@@ -159,10 +159,16 @@ class MPClient(EngineCoreClient):
atexit.register(self.shutdown)
def shutdown(self):
# During final garbage collection in process shutdown, atexit may be
# None.
if atexit:
# in case shutdown gets called via __del__ first
atexit.unregister(self.shutdown)
# Shut down the zmq context.
self.ctx.destroy(linger=0)
if hasattr(self, "proc_handle"):
if hasattr(self, "proc_handle") and self.proc_handle:
# Shutdown the process if needed.
if self.proc_handle.proc.is_alive():
self.proc_handle.proc.terminate()
@@ -178,8 +184,9 @@ class MPClient(EngineCoreClient):
]
for ipc_socket in ipc_sockets:
socket_file = ipc_socket.replace("ipc://", "")
if os.path.exists(socket_file):
if os and os.path.exists(socket_file):
os.remove(socket_file)
self.proc_handle = None
def __del__(self):
self.shutdown()
+7
View File
@@ -196,3 +196,10 @@ class LLMEngine:
f"found type: {type(tokenizer_group)}")
return tokenizer_group
def __del__(self):
self.shutdown()
def shutdown(self):
if engine_core := getattr(self, "engine_core", None):
engine_core.shutdown()
+9 -1
View File
@@ -163,6 +163,10 @@ class MultiprocExecutor:
termination and kill signals if needed."""
def wait_for_termination(procs, timeout):
if not time:
# If we are in late stage shutdown, the interpreter may replace
# `time` with `None`.
return all(not proc.is_alive() for proc in procs)
start_time = time.time()
while time.time() - start_time < timeout:
if all(not proc.is_alive() for proc in procs):
@@ -187,10 +191,14 @@ class MultiprocExecutor:
for w in self.workers:
# Remove the zmq ipc socket file
socket_path = w.ready_path.replace("ipc://", "")
if os.path.exists(socket_path):
if os and os.path.exists(socket_path):
os.remove(socket_path)
def shutdown(self):
if atexit:
# in case shutdown was called explicitly, we don't need to call it
# again
atexit.unregister(self.shutdown)
"""Properly shut down the executor and its workers"""
if (hasattr(self, 'workers') and self.workers is not None):
for w in self.workers: #TODO: not sure if needed