Gpu Cuda
Runs CUDA compute workloads inside a smolvm microVM against a real host NVIDIA GPU...
Install
npx skills add smol-machines/smolvm --skill gpu-cuda -g -a claude-code -y- Repository
- smol-machines/smolvm6.9k
- Category
- Media
- Updated
- 3d ago
- License
- Apache-2.0
- Context cost
- Mediumabout 3,000 tokens
- Runs scripts
- Yes, review before use
What it does
Runs CUDA compute workloads inside a smolvm microVM against a real host NVIDIA GPU, using smolvm's --cuda API remoting. Use when a workload in a machine needs a GPU; when nvidia-smi or /dev/nvidia* is missing inside a --cuda guest; when a CUDA program in a machine exits zero but seems not to touch the device; when the shim will not load in an Alpine image; or when checking whether CUDA is available on a given platform, including Windows. Do not use it for Vulkan graphics (--gpu), which is a separate feature, in docs/gpu.md, and do not expect it on a Mac, which has no CUDA driver.
CUDA inside a machine
Verified on smolvm v1.22.2 against an NVIDIA GeForce RTX 4050 (driver 32.0.15.6626, Windows x86_64), 2026-10-03, and on v1.14.6 against an NVIDIA A10 (driver 580.105.08, Linux x86_64, kernel 6.8.0-1046-nvidia). Done means a program in the VM opens the device, creates a context, and moves data to and from it. The Linux runs used the scripts of their date; this version's preflight and cleanup scripts ran on Linux aarch64 on v1.22.2 on 2026-10-03.
The GPU path was last run end to end on Linux x86_64 on v1.14.6, on an A10: the preflight, the probe against two glibc images, all three eval prompts and the cleanup. Windows was re-run on v1.22.2 against an RTX 4050: the probe, the two absences, the shim in a plain
alpineguest, and a vector add over 16777216 elements; seereferences/windows.md. The macOS arm64 and Linux aarch64 hosts have no NVIDIA hardware, so what runs there is the no-GPU answer, on v1.22.2 on macOS and v1.18.2 on Linux aarch64: the preflight reportsgpu_present=nowithout starting anything, and--cudareaches a CPU emulation device, which the probe names rather than passing. "What was not run" lists every remaining step.
How this works, and why the checks are shaped this way
The guest gets no NVIDIA driver and no /dev/nvidia*. A compatibility libcuda.so.1 is
injected at /opt/smolvm-cuda inside the guest and forwards CUDA driver calls over vsock to a host
daemon that owns the device. Nothing is needed on the host beyond a working NVIDIA driver: no extra
packages, no container toolkit, no device plugin.
Because the API is remoted rather than passed through, a program can start, link against
libcuda.so.1 and exit zero without a GPU ever being reached. Assert a device name and a
transferred result, never an exit code.
Procedure
1. Preflight.
scripts/preflight.sh
