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
[BugFix][Core] Fix error when enable async-scheduling in multi-node env (#25887)
Signed-off-by: Lehua Ding <lehuading@tencent.com> Signed-off-by: Lehua Ding <lehuading@qq.com> Co-authored-by: Benjamin Chislett <chislett.ben@gmail.com>
Lehua Ding committed
Oct 18, 2025 at 06:16 UTC
6367bde739dc3cd558f22d8db2021098c6435e31
1 file changed
+9
-7
vllm/engine/arg_utils.py
+9
-7
@@ -1440,13 +1440,6 @@ class EngineArgs:
1440
)
1441
1442
if self.async_scheduling:
1443
- # Async scheduling does not work with the uniprocess backend.
1444
- if self.distributed_executor_backend is None:
1445
- self.distributed_executor_backend = "mp"
1446
- logger.info(
1447
- "Defaulting to mp-based distributed executor "
1448
- "backend for async scheduling."
1449
- )
1443
if self.pipeline_parallel_size > 1:
1444
raise ValueError(
1445
"Async scheduling is not supported with pipeline-parallel-size > 1."
@@ -1503,6 +1496,15 @@ class EngineArgs:
1496
_api_process_rank=self._api_process_rank,
1497
)
1498
1499
+ if self.async_scheduling and (
1500
+ parallel_config.distributed_executor_backend not in ("mp", "uni")
1501
+ ):
1502
+ raise ValueError(
1503
+ "Currently, async scheduling only supports `mp` or `uni` "
1504
+ "distributed executor backend, but you choose "
1505
+ f"`{parallel_config.distributed_executor_backend}`."
1506
+ )
1507
+
1508
speculative_config = self.create_speculative_config(
1509
target_model_config=model_config,
1510
target_parallel_config=parallel_config,