A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 8, 01:15 UTC
[Core] Fix state names in pause_scheduler() (#34840)
Signed-off-by: Mark McLoughlin <markmc@redhat.com>
Mark McLoughlin committed
Feb 20, 2026 at 01:21 UTC
76df6072ff4829980ad71764191fc970a873275a
1 file changed
+9
-9
vllm/v1/engine/core.py
+9
-9
@@ -1370,15 +1370,15 @@ class EngineCoreProc(EngineCore):
1370
) -> Future | None:
1371
"""Pause generation; behavior depends on mode.
1372
1373
- All pause states queue new adds. PAUSE_ABORT and PAUSE_KEEP skip step();
1374
- PAUSE_WAIT allows step() so in-flight requests can drain.
1375
-
1376
- - ``abort``: Set PAUSE_ABORT, abort all requests, wait for abort
1377
- outputs to be sent (when running with output_queue), clear caches,
1378
- then complete the returned Future.
1379
- - ``wait``: Set PAUSE_WAIT (queue adds, keep stepping); when drained,
1380
- set PAUSE_KEEP, clear caches, complete the returned Future.
1381
- - ``keep``: Set PAUSE_KEEP; return a Future that completes when the
1373
+ All pause modes queue new adds -- "abort" and "keep" skip step();
1374
+ "wait" allows step() so in-flight requests can drain.
1375
+
1376
+ - ``abort``: Set PAUSED_NEW, abort all requests, wait for abort
1377
+ outputs to be sent (when running with output_queue), optionally
1378
+ clear caches, then complete the returned Future.
1379
+ - ``wait``: Set PAUSED_NEW (queue adds, keep stepping); when drained,
1380
+ optionally clear caches, then complete the returned Future.
1381
+ - ``keep``: Set PAUSED_ALL; return a Future that completes when the
1382
output queue is empty.
1383
"""
1384
if mode not in ("keep", "abort", "wait"):