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] Fix DBO hang (#25625)
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Lucas Wilkinson committed
Sep 25, 2025 at 13:04 UTC
13cc7f5370ed5d629028143311e95ac6646819cb
1 file changed
+2
-1
vllm/v1/worker/gpu_ubatch_wrapper.py
+2
-1
@@ -366,7 +366,8 @@ class UBatchWrapper:
366
cudagraph_runtime_mode=CUDAGraphMode.NONE)
367
with self.sm_control:
368
return self._capture_ubatches(ubatch_metadata, self.model)
369
- elif num_tokens in self.cudagraphs:
369
+ elif num_tokens in self.cudagraphs \
370
+ and cudagraph_runtime_mode is CUDAGraphMode.FULL:
371
cudagraph_metadata = self.cudagraphs[num_tokens]
372
cudagraph_metadata.cudagraph.replay()
373
return cudagraph_metadata.outputs