All pages
Docs · Start hereMarkdown
Install
Public binaries are coming soon. To join early access, email niko@041.io.
What a machine needs
The binary is self-contained. Kernels are compiled at run time from source inside it, so the machine that trains needs no Python and no Rust toolchain.
| to | the machine needs |
|---|---|
check, explain, fmt and every authoring verb | nothing; any machine the binary runs on |
run --device cpu, the interpreter | nothing; toy sizes only, see devices |
run --device cuda | a CUDA 13 driver (580 series or newer) and an Ampere or newer GPU: A100, L4, H100 |
The full CUDA requirements are in devices.
Check the machine
sexpgpu doctor
doctor prints one line per thing (the binary, the CUDA build, the GPU and
its driver and compute-capability floor, the environment, credentials, disk)
and exits 1 for the configurations that stop a run before its first step.
A failed floor is printed but does not change the exit code; read the
floor line. See doctor.
Next: the quickstart.