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 8, 07:30 UTC
[Misc] vLLM's --enforce-eager should turn off compile and cudagraphs only (#34523)
Signed-off-by: Richard Zou <zou3519@gmail.com>
Richard Zou committed
Feb 13, 2026 at 12:52 UTC
87789c836422cf3b666ddf3eca9ede8e03f735ee
1 file changed
+7
-7
vllm/config/vllm.py
+7
-7
@@ -728,13 +728,13 @@ class VllmConfig:
728
"precision for chunked prefill triton kernels."
729
)
730
731
- if (
732
- self.optimization_level > OptimizationLevel.O0
733
- and self.model_config is not None
734
- and self.model_config.enforce_eager
735
- ):
736
- logger.warning("Enforce eager set, overriding optimization level to -O0")
737
- self.optimization_level = OptimizationLevel.O0
731
+ if self.model_config is not None and self.model_config.enforce_eager:
732
+ logger.warning(
733
+ "Enforce eager set, disabling torch.compile and CUDAGraphs. "
734
+ "This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none"
735
+ )
736
+ self.compilation_config.mode = CompilationMode.NONE
737
+ self.compilation_config.cudagraph_mode = CUDAGraphMode.NONE
738
739
if self.compilation_config.backend == "eager" or (
740
self.compilation_config.mode is not None