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 11, 12:30 UTC
[Log] Skip piecewise cudagraph warn when using full cudagraph (#30657)
Signed-off-by: Boyuan Feng <boyuan@meta.com>
Boyuan Feng committed
Dec 14, 2025 at 18:49 UTC
917fdae5b2eccf0e7b6f2d4ae67132d13d13580c
1 file changed
+7
-3
vllm/config/compilation.py
+7
-3
@@ -932,9 +932,13 @@ class CompilationConfig:
932
self.splitting_ops = list(self._attention_ops)
933
added_default_splitting_ops = True
934
elif len(self.splitting_ops) == 0:
935
- logger.warning_once(
936
- "Using piecewise compilation with empty splitting_ops"
937
- )
935
+ if (
936
+ self.cudagraph_mode == CUDAGraphMode.PIECEWISE
937
+ or self.cudagraph_mode == CUDAGraphMode.FULL_AND_PIECEWISE
938
+ ):
939
+ logger.warning_once(
940
+ "Using piecewise compilation with empty splitting_ops"
941
+ )
942
if self.cudagraph_mode == CUDAGraphMode.PIECEWISE:
943
logger.warning_once(
944
"Piecewise compilation with empty splitting_ops do not"