[sleep functionality] code refactor about sleep/wake_up (#50431)
Signed-off-by: Andy Xie <andy.xning@gmail.com>
Ning Xie committed
Aug 25, 2026 at 15:45 UTC
23310bc3de530fd90ea060fcc1ef7f58501ee859
3 files changed
+4
-7
vllm/entrypoints/serve/dev/sleep/api_router.py
-4
@@ -24,8 +24,6 @@ async def sleep(raw_request: Request):
24
level = raw_request.query_params.get("level", "1")
25
mode = raw_request.query_params.get("mode", "abort")
26
await engine_client(raw_request).sleep(int(level), mode)
27
- # FIXME: in v0 with frontend multiprocessing, the sleep command
28
- # is sent but does not finish yet when we return a response.
27
return Response(status_code=200)
28
29
@@ -37,8 +35,6 @@ async def wake_up(raw_request: Request):
35
tags = None
36
logger.info("wake up the engine with tags: %s", tags)
37
await engine_client(raw_request).wake_up(tags)
40
- # FIXME: in v0 with frontend multiprocessing, the wake-up command
41
- # is sent but does not finish yet when we return a response.
38
return Response(status_code=200)
39
40
vllm/v1/core/sched/scheduler.py
+1
@@ -2476,6 +2476,7 @@ class Scheduler(SchedulerInterface):
2476
return self._pause_state
2477
2478
def set_pause_state(self, pause_state: PauseState) -> None:
2479
+ logger.info("setting pause state to %s", pause_state.name)
2480
self._pause_state = pause_state
2481
2482
def _free_request_blocks(self, request: Request):
vllm/v1/engine/core.py
+3
-3
@@ -15,7 +15,7 @@ from functools import partial
15
from inspect import isclass, signature
16
from logging import DEBUG
17
from multiprocessing.queues import Queue
18
-from typing import Any, TypeVar, cast
18
+from typing import Any, TypeVar, cast, get_args
19
20
import msgspec
21
import zmq
@@ -855,7 +855,7 @@ class EngineCore:
855
- ``keep``: Set PAUSED_ALL; return a Future that completes when the
856
output queue is empty.
857
"""
858
- if mode not in ("keep", "abort", "wait"):
858
+ if mode not in get_args(PauseMode):
859
raise ValueError(f"Invalid pause mode: {mode}")
860
if mode == "wait":
861
raise ValueError("'wait' mode can't be used in inproc-engine mode")
@@ -1933,7 +1933,7 @@ class EngineCoreProc(EngineCore):
1933
- ``keep``: Set PAUSED_ALL; return a Future that completes when the
1934
output queue is empty.
1935
"""
1936
- if mode not in ("keep", "abort", "wait"):
1936
+ if mode not in get_args(PauseMode):
1937
raise ValueError(f"Invalid pause mode: {mode}")
1938
1939
def engine_idle_callback(engine: "EngineCoreProc", future: Future[Any]) -> None: