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 5, 06:00 UTC
[Pooling] Disable async scheduling by default for pooling models (#39592)
Signed-off-by: Nick Hill <nickhill123@gmail.com>
Nick Hill committed
Apr 12, 2026 at 00:23 UTC
ee3c0c83db824d942494faec05820a56acbcda1a
1 file changed
+10
vllm/config/vllm.py
+10
@@ -764,6 +764,16 @@ class VllmConfig:
764
elif self.scheduler_config.async_scheduling is None:
765
# Enable async scheduling unless there is an incompatible option.
766
if (
767
+ self.model_config is not None
768
+ and self.model_config.runner_type == "pooling"
769
+ ):
770
+ # The current implementation of asynchronous scheduling negatively
771
+ # impacts performance of pooling models, so we disable by default.
772
+ logger.debug(
773
+ "Disabling asynchronous scheduling by default for pooling model."
774
+ )
775
+ self.scheduler_config.async_scheduling = False
776
+ elif (
777
self.speculative_config is not None
778
and self.speculative_config.method not in get_args(EagleModelTypes)
779
and self.speculative_config.method not in get_args(NgramGPUTypes)