NVIDIA - auto install WSLg DirectGPU between Windows and Ubuntu WSL

[cyberluke]

Basically the extension should be called Ultimate WSL2 & WSLg Automasturbator

• install RAPIDS should be enough (it will install CUDA)

• no Linux Geforce driver! important

• use WSLg and DirectGPU to link libcuda.so (should be automatic)

• provide simple dropdown menu to select cuda version - ideally run nvidia-smi on Windows, check CUDA version (ie. 12.8) and run RAPIDS install inside Ubuntu WSL with that CUDA 12.8 version

• allow to choose RAPIDS install using PIP, Conda or Docker container (recommended)

• in case of errors, provide a hint where to download Windows CUDA and drivers and how to set up env PATH in case of version mismatch, but more CUDA versions avail on Win

• nitpicking guide to make it running (AI cannot do it)

• show how you can run GUI, for example ComfyUI in Docker Linux and render only that window in Windows using WSLg (before X410 project was doing it using Socks streaming from X11)

• automatize creating Python venv per project inside container

• utilize all VRAM

• support Intel NPU

• support iGPU (Intel Xe or Arc) together with Geforce RTX

• utilize all CPUs (i9 Ultra Core 285K has 24 processors, but WSL always hardcode 32 and that does not work)

• let dynamically allocate RAM (for example machine have 128GB, start with 50% - 64GB)

• know how to auto update RAPIDS (will ask for user permission each update because it can break everything) + update PyTorch CUDA in each venv created inside container

• can provide NANOTRIK.AI custom OSS OpenAI gateway API including OpenVINO, vLLM or SGLang, LMStudio, Ollama, etc (contact me)