// nix2gpu //

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

nixos containers for cost-effective and capable gpu compute. vast.ai is the first target, there will be many more.

// init // what this is

a nix-based container runtime that makes distributed gpu compute actually work. reproducible environments, tailscale networking, CUDA 12.8. everything you need to turn random gpus into a coherent compute cluster.

vast.ai is just the beginning. any gpu, anywhere, real init system, same environment.

// tree // architecture

├── checks       # flake checks run on this repo
├── config       # avalible `nix2gpu` config options
├── dev          # `flake-parts` config for development of this repo
├── examples     # working `nix2gpu` examples
├── services     # extra modular service definitions
└── modules      # implementation of `nix2gpu`

clean separation of concerns. the flake stays minimal, modules do the work.

// quick // start

create a nix2gpu container:

perSystem.nix2gpu.sample = { };

# build
nix build .#sample

# push to registry
nix run .#sample.copyToGIthub

# run locally (docker or podman)
nix run .#sample.copyToContainerRuntime
nix run .#sample.shell

// manifest // what's inside

• cuda 12.8 with cudnn, nccl, cublas - the whole show
• tailscale for seamless and secure networking across a heterogeneous fleet
• development tools - gcc, python, uv, patchelf
• modern shell - tmux, starship, atuin, ripgrep, fzf, the usual suspects
• nix - because docker/OCI is a reasonable deployment target but an unreasonable build system

// deployment

push to ghcr.io, point vast.ai at it:

ghcr.io/weyl-ai/nix2gpu:20250914-142437

or run your own registry. we don't care.

// configuration

# optional
ROOT_PASSWORD=...
SSH_PUBLIC_KEYS=...
TAILSCALE_AUTHKEY=...

passwordless root by default. we're already inside the machine.

// technical details

vast.ai uses debian library paths (/lib/x86_64-linux-gnu). their nvidia-container-runtime injects driver libraries at container start. we handle this gracefully:

• wait for library injection
• patch nvidia-smi with nix interpreter
• regenerate ldconfig cache
• seamless thereafter

it just works. deterministically.

// development

nix develop      # enter shell
nix build        # build image
nix run .#<app>  # run apps

// philosophy

cloud vendors want you to think gpu compute is complicated. it's not. it's just linux with extra libraries.

this project is an existence proof that complexity is optional.

// contributing

send patches. fix things. make it better.

the code is MIT. use it however you want.

// notes // support

built at fleek. some testing in production. works on our machines, will expect it will work on yours. if not, we'll fix it.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━