Skip to content

SomeoneSerge/pkgs

Folders and files

NameName
Last commit message
Last commit date
Jan 31, 2025
Sep 26, 2023
Oct 22, 2023
Jan 19, 2022
Jan 9, 2024
Mar 4, 2024
Jul 27, 2023
Oct 1, 2023
Sep 26, 2023
Jan 19, 2022
Jan 19, 2022
Aug 23, 2023
Nov 20, 2023
Nov 20, 2023
May 13, 2022
Oct 23, 2024
Jan 25, 2024
Dec 3, 2023
Oct 22, 2023
Apr 7, 2022
Apr 3, 2023

Repository files navigation

nix-build nix-update nix flake update

Cachix Cache

Deploying the undeployable with Nix

Sci-Comp and Deep Learning projects are notoriously hard to build and deploy: even to satisfy the build-time dependencies people inevitably fall back to building things in Docker (build-time isolation). And when building things in Docker, they would some times fetch dependencies through conda. And then they would take the stock they've built through sweat and blood, and run it in containers (run-time isolation). And that is just the beginning: later, for example, they'll seek for ways to establish a communication between the container and the host system's X-server (that they've put so much effort in isolating from).

When they have gone through all these pains and struggles, and had cheered at a running visualizer from some old baseline project, in a month or two they should probably find that, in spite of all effort, their hard-won trophey won't acknowledge their CUDA installation anymore, and nobody knows why.

Now, have you really considered if there ever was a need for Docker or Kubernetes?

Enter Nix (for building things) and NixOS (for setting up the runtime environment, notably /run/opengl-driver/lib):

pkgs-instant-ngp-1.mp4

Transcript:

nix build github:SomeoneSerge/pkgs/unfree#instant-ngp.data
nix build github:SomeoneSerge/pkgs/unfree#instant-ngp.configs
nix run github:SomeoneSerge/pkgs/unfree#instant-ngp -- --scene ./result-data/nerf/fox/ --config ./result-configs/nerf/base.json

DISCLAIMER:

  • The video demonstrates https://github.com/NVlabs/instant-ngp

  • /run/opengl-driver/lib/

    Nix deploys its executables and libraries with absolute paths to concrete versions of their dependencies that are guaranteed to work (e.g. through Runpath in ELF headers on Linux). However, this doesn't apply to hardware-specific dependencies, like graphics and compute drivers. For nix run to work with OpenGL or CUDA-enabled applications, one needs to symlink their hardware-dependent drivers in /run/opengl-driver/lib. NixOS does that automatically when hardware.opengl.enable = true

  • NixGL

    When running on Ubuntu&c one may face an issue that their graphics driver links against a revision libc older than the one used by this repo. In these cases a temporary workaround could be NixGL.