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 14, 05:30 UTC
[Core] Enable async scheduling for external_launcher mode (#27394)
Signed-off-by: 22quinn <33176974+22quinn@users.noreply.github.com> Co-authored-by: Zhuohan Li <zhuohan123@gmail.com>
22quinn committed
Oct 28, 2025 at 06:52 UTC
2abbd351ef6436860cb3c64f66b1452c1d941fe4
2 files changed
+5
-6
vllm/engine/arg_utils.py
+4
-3
@@ -1553,11 +1553,12 @@ class EngineArgs:
1553
)
1554
1555
if self.async_scheduling and (
1556
- parallel_config.distributed_executor_backend not in ("mp", "uni")
1556
+ parallel_config.distributed_executor_backend
1557
+ not in ("mp", "uni", "external_launcher")
1558
):
1559
raise ValueError(
1559
- "Currently, async scheduling only supports `mp` or `uni` "
1560
- "distributed executor backend, but you choose "
1560
+ "Currently, async scheduling only supports `mp`, `uni` or "
1561
+ "`external_launcher` distributed executor backend, but you choose "
1562
f"`{parallel_config.distributed_executor_backend}`."
1563
)
1564
vllm/v1/engine/llm_engine.py
+1
-3
@@ -306,9 +306,7 @@ class LLMEngine:
306
self.engine_core.abort_requests(processed_outputs.reqs_to_abort)
307
308
# 4) Record stats
309
- if self.logger_manager is not None:
310
- assert outputs.scheduler_stats is not None
311
-
309
+ if self.logger_manager is not None and outputs.scheduler_stats is not None:
310
self.logger_manager.record(
311
scheduler_stats=outputs.scheduler_stats,
312
iteration_stats=iteration_stats,