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 9, 07:30 UTC
[BugFix] Disable async scheduling for Mamba prefix caching (#33352)
Signed-off-by: huanghaoyan.hhy <huanghaoyan.hhy@alibaba-inc.com>
Harry Huang committed
Jan 30, 2026 at 12:40 UTC
ec51831a22cbb434646a5d8219c694ab15dbc4cb
1 file changed
+12
vllm/config/vllm.py
+12
@@ -619,6 +619,11 @@ class VllmConfig:
619
"`external_launcher` distributed executor backend, but you chose "
620
f"`{executor_backend}`."
621
)
622
+ if self.cache_config.mamba_cache_mode != "none":
623
+ raise ValueError(
624
+ "Currently, async scheduling is not compatible with "
625
+ "prefix caching for Mamba models."
626
+ )
627
elif self.scheduler_config.async_scheduling is None:
628
# Enable async scheduling unless there is an incompatible option.
629
if (
@@ -651,6 +656,13 @@ class VllmConfig:
656
scope="local",
657
)
658
self.scheduler_config.async_scheduling = False
659
+ elif self.cache_config.mamba_cache_mode != "none":
660
+ logger.warning_once(
661
+ "Async scheduling is not compatible with "
662
+ "prefix caching for Mamba models and will be disabled.",
663
+ scope="local",
664
+ )
665
+ self.scheduler_config.async_scheduling = False
666
else:
667
self.scheduler_config.async_scheduling = True
668