GPU detection, a working CUDA computation, and execution on the intended backend are separate checks. The reports below test those checks on specific Windows/WSL2 configurations. Each links to its commands and recorded output.
| Check you need | Tested report | What the result covers |
|---|---|---|
CUDA availability is False | RTX 3060 Laptop: four reproduced causes | Driver/wheel mismatch, CPU wheel, GPU visibility, and interpreter selection; recovery checked with a CUDA matrix multiplication |
| CUDA is available but an operation fails | RTX 5060 Ti: PyTorch 2.13 cu126, cu130, and cu132 | A fixed matrix of tensor creation, copies, pointwise operations, dot, and GEMM; separate compiler prerequisite checks |
| A requested CUDA feature runs slowly | RTX 5060 Ti: llama.cpp mixed KV cache | Scheduler traces for mixed K/V Flash Attention at one pinned commit, plus measurements with the build flag changed |
Match the environment before the recommendation
The RTX 3060 Laptop report tested a driver mismatch and a compatible wheel on that host. The RTX 5060 Ti report tested a different GPU, driver, and kernel support. A wheel recommendation from the first experiment is not a compatibility verdict for the second GPU.
Record your GPU model, Windows driver, WSL version, Python interpreter, PyTorch version, and torch.version.cuda. Use the diagnostic commands in the matching report to compare the actual failure stage. A displayed CUDA version alone does not establish that the operation you need will run.
What “passed” means here
A pass applies to the commands and checks in the report. The PyTorch matrix is a compatibility probe, not a performance ranking or proof that every model works. The llama.cpp diagnostic proves scheduler placement for the inspected nodes at its pinned commit; registered CUDA buffers alone would not prove that placement.
For measured speed and memory results, see Benchmarks. For disk and cache issues encountered before CUDA setup, see Fixes.