[Feat] Allow custom naming of vLLM processes (#21445)

Signed-off-by: chaunceyjiang <chaunceyjiang@gmail.com>
This commit is contained in:
Chauncey
2025-07-24 03:15:23 -07:00
committed by GitHub
parent 73e3949d07
commit 6da0078523
10 changed files with 46 additions and 17 deletions
+6 -5
View File
@@ -13,7 +13,8 @@ from vllm.logger import init_logger
from vllm.utils import get_mp_context, make_zmq_socket
from vllm.v1.engine import EngineCoreOutputs, EngineCoreRequestType
from vllm.v1.serial_utils import MsgpackDecoder
from vllm.v1.utils import get_engine_client_zmq_addr, shutdown
from vllm.v1.utils import (bind_process_name, get_engine_client_zmq_addr,
shutdown)
logger = init_logger(__name__)
@@ -79,7 +80,7 @@ class DPCoordinator:
context = get_mp_context()
self.proc: multiprocessing.Process = context.Process(
target=CoordinatorProc.run_coordinator,
target=DPCoordinatorProc.run_coordinator,
name="VLLM_DP_Coordinator",
kwargs={
"engine_count": parallel_config.data_parallel_size,
@@ -113,12 +114,12 @@ class EngineState:
self.request_counts = [0, 0] # [waiting, running]
class CoordinatorProc:
class DPCoordinatorProc:
def __init__(self,
engine_count: int,
min_stats_update_interval_ms: int = 100):
bind_process_name(self.__class__.__name__)
self.ctx = zmq.Context()
self.engines = [EngineState() for _ in range(engine_count)]
@@ -137,7 +138,7 @@ class CoordinatorProc:
back_publish_address: str,
min_stats_update_interval_ms: int = 100,
):
coordinator = CoordinatorProc(
coordinator = DPCoordinatorProc(
engine_count=engine_count,
min_stats_update_interval_ms=min_stats_update_interval_ms)
try:
+3 -1
View File
@@ -25,7 +25,8 @@ from vllm.logging_utils.dump_input import dump_engine_exception
from vllm.lora.request import LoRARequest
from vllm.transformers_utils.config import (
maybe_register_config_serialize_by_value)
from vllm.utils import make_zmq_socket, resolve_obj_by_qualname
from vllm.utils import (bind_process_name, make_zmq_socket,
resolve_obj_by_qualname)
from vllm.v1.core.kv_cache_utils import (get_kv_cache_config,
unify_kv_cache_configs)
from vllm.v1.core.sched.interface import SchedulerInterface
@@ -411,6 +412,7 @@ class EngineCoreProc(EngineCore):
client_handshake_address: Optional[str] = None,
engine_index: int = 0,
):
bind_process_name(self.__class__.__name__, f"{engine_index}")
self.input_queue = queue.Queue[tuple[EngineCoreRequestType, Any]]()
self.output_queue = queue.Queue[Union[tuple[int, EngineCoreOutputs],
bytes]]()