biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Perf] Change default CUDAGraphMode from PIECEWISE to FULL_AND_PIECEWISE (#25444)

Browse Source 
Signed-off-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Michael Goin
2025-09-23 15:29:26 -04:00
committed by GitHub
parent 63400259d0
commit 24fab45d96
3 changed files with 32 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
vllm/config/__init__.py
View File

@@ -509,8 +509,15 @@ class VllmConfig:
if self.compilation_config.cudagraph_mode is None:
if envs.VLLM_USE_V1 and self.compilation_config.level \
== CompilationLevel.PIECEWISE:
# default to full and piecewise for most models
self.compilation_config.cudagraph_mode = \
CUDAGraphMode.PIECEWISE
CUDAGraphMode.FULL_AND_PIECEWISE
# pooling model does not support full cudagraphs
if self.model_config is not None and \
self.model_config.pooler_config is not None:
self.compilation_config.cudagraph_mode = \
CUDAGraphMode.PIECEWISE
else:
self.compilation_config.cudagraph_mode = CUDAGraphMode.NONE