GPU / VRAM («Execute unsloth code on a computer without gpu but finetuning of a cluster with nvidia gpus»)

1. Match driver ↔ CUDA toolkit ↔ wheel (PyTorch +cu128 vs system nvcc can disagree). python -c "import torch; print(torch.cuda.is_available(), torch.version.cuda)".
2. OOM: drop batch, n_ctx / context, quantization, --parallel. WSL2 also needs .wslconfig memory= or the VM eats the host.
3. “GPU not used”: process might be on CPU because the build is CPU-only, or CUDA_VISIBLE_DEVICES is empty.

Add: GPU model, VRAM, driver, and whether this is native Windows, WSL, or Docker.