This example is for hosts where:
nvidia-smiworks inside WSL/dev/dxgexists- but Vulkan may or may not be complete
Use the wrapper script and let it choose:
cudawhen CUDA userland indicators are presentvulkanwhen Vulkan tooling is presentfalseotherwise
nvidia-smi
ls -l /dev/dxg
QMD_LLAMA_GPU=cuda qmd status
QMD_LLAMA_GPU=vulkan qmd status
QMD_LLAMA_GPU=false qmd statusYou may have GPU passthrough visible but still miss Vulkan build tools:
libvulkan-devvulkan-toolsglslcglslang-tools
In that case, a stable CPU fallback is better than repeated failed auto-detection.