diff --git a/vllm/platforms/cuda.py b/vllm/platforms/cuda.py index 4059a7e4f5..f9304dfb2a 100644 --- a/vllm/platforms/cuda.py +++ b/vllm/platforms/cuda.py @@ -290,7 +290,8 @@ class CudaPlatformBase(Platform): # kernel with limited pinned memory support for CUDA. version = _get_wsl_kernel_version() if version is None or version < (4, 19, 121): - logger.warning_once( + # warning_once() causes a circular import on WSL, see #48397. + logger.warning( "Using 'pin_memory=False' as WSL is detected and the " "WSL2 kernel version is below 4.19.121. This may slow " "down performance. Please run `wsl --update`." diff --git a/vllm/platforms/interface.py b/vllm/platforms/interface.py index 7cf009c807..07dce7d138 100644 --- a/vllm/platforms/interface.py +++ b/vllm/platforms/interface.py @@ -991,7 +991,8 @@ class Platform: # Pinned memory support under WSL depends on the vendor and driver # version. Conservative default: return False. Platform subclasses # that can verify support (e.g. CudaPlatformBase) override this. - logger.warning_once( + # warning_once() causes a circular import on WSL, see #48397. + logger.warning( "Using 'pin_memory=False' as WSL is detected. " "This may slow down performance." )