!eWOErHSaiddIbsUNsJ:nixos.org

NixOS CUDA

325 Members
CUDA packages maintenance and support in nixpkgs | https://github.com/orgs/NixOS/projects/27/ | https://nixos.org/manual/nixpkgs/unstable/#cuda64 Servers

Load older messages


SenderMessageTime
10 May 2026
@hexa:lossy.networkhexaand they have their own branches and rules how they propagate from branch to branch18:15:18
@neobrain:matrix.orgneobrainyeah I already tried changing my nixpkgs input to follow flox's, so who knows what's wrong 🤷18:16:13
@tomberek:matrix.orgtomberekThis has been a long-standing problem, how to track what is cached and the pins. The underlying issue is that it makes the caches a critical component of the publishing process, not just an optimization. One approach is to encode this into "fake derivations" that promise to be substitutable (eg: via fetchClosure or similar mechanisms). The issue with your attempt on the flox cache is using the wrong nixpkgs. Are you tracking one of the "stable"/"unstable" branches at https://github.com/flox/nixpkgs/tree/stable ?18:23:30
@glepage:matrix.orgGaétan Lepage We (nixos-cuda.org) build all nixpkgs packages with cudaSupport = true (as long as the hash differs from the non-cudaSupport version).
Our jobsets are publicly available here.
19:26:07
@glepage:matrix.orgGaétan Lepage

Having said that, the cache.nixos-cuda.org binary cache is not meant to be consumed by the general public. NVIDIA restricts how certains packages should not be distributed after being patched. This cache is a development resource for the cuda team members.
Flox is the only nix-related entity allowed to redistribute those packages.

https://github.com/nixos-cuda/infra?tab=readme-ov-file#binary-cache

19:28:17
@glepage:matrix.orgGaétan Lepage *

Having said that, the cache.nixos-cuda.org binary cache is not meant to be consumed by the general public. NVIDIA restricts the redistribution of certain packages after having been patched. This cache is a development resource for the cuda team members.
Flox is the only nix-related entity allowed to redistribute those packages.

https://github.com/nixos-cuda/infra?tab=readme-ov-file#binary-cache

19:29:14
@glepage:matrix.orgGaétan Lepage *

Having said that, the cache.nixos-cuda.org binary cache is not meant to be consumed by the general public. NVIDIA restricts the redistribution of certain packages after having been patched. This cache is a development resource for the nix cuda team members.
Flox is the only nix-related entity allowed to redistribute those packages.

https://github.com/nixos-cuda/infra?tab=readme-ov-file#binary-cache

19:29:55
@neobrain:matrix.orgneobrain yes, I'm using nixpkgs.url = "github:flox/nixpkgs/unstable" in my flake inputs 20:24:10
@neobrain:matrix.orgneobrainI just remembered that nixos-cuda.org doesn't do aarch64 yet, so I can't give it a try after all :(20:25:01
@hexa:lossy.networkhexadoes flox do aarch64-linux??/20:48:54
@hexa:lossy.networkhexaRedacted or Malformed Event20:48:56
@hexa:lossy.networkhexaI'd say right now the problem is that renting aarch64-linux for cheap is not really possible20:49:27
@neobrain:matrix.orgneobrainAh! nixos-dgx-spark overrides nixpkgs.config.cudaCapabilities in an attempt to improve compile times, but then the side effect of it is that the slim configuration isn't cached by flox. Resetting the variable gives me cache hits again20:50:21
@neobrain:matrix.orgneobrainYup they do20:50:34
@neobrain:matrix.orgneobrainyeah that makes sense :(20:51:39
@sporeray:matrix.orgRobbie Buxton
In reply to @hexa:lossy.network
I'd say right now the problem is that renting aarch64-linux for cheap is not really possible
I don’t think it’s too bad in hyper scalers (graviton isn’t bad) not sure about other places tho
20:54:26
@hexa:lossy.networkhexayeah, but cheap :D20:57:14
@sporeray:matrix.orgRobbie Buxton Last time I checked graviton was similar if not cheaper 20:58:00
@neobrain:matrix.orgneobrainI'm guessing self-hosting has been considered but isn't an option? nvidia has a pretty generous marketing budget for free hardware21:01:09
@glepage:matrix.orgGaétan Lepage aarch64-linux support is definitely on our roadmap. For now, the critical blocker is getting access to relevant hardware. I'll give updates in this channel as soon as we get new opportunities in this regard. 21:01:46
@mjolnir:nixos.orgNixOS Moderation Botchanged room power levels.21:07:38
11 May 2026
@glepage:matrix.orgGaétan Lepage Looking for an approval on https://github.com/NixOS/nixpkgs/pull/517764 (cudaPackages.cudnn update).
connor (burnt/out) (UTC-8) maybe?
15:52:02
@smudgebun:matrix.orgsmudge (she/her)Redacted or Malformed Event19:34:06
@smudgebun:matrix.orgsmudge (she/her) should the nixos wiki cuda page's cache section be corrected? it recommends the cache.nixos-cuda.org cache and doesn't mention others which lead me to assume it was the one for public use. is nix-community.cachix.org (per this announcement) the one that should be primarily recommended on the wiki? as well as information about using flox's nixpkgs repo and its cache? Or, if the nix-community.cachix.org isn't actually allowed to distribute should flox be the primary suggested cache?

also should the information for cache.nixos-cuda.org be retained, just with a proper disclaimer that it is for internal use only, or how would you like that to be handled?
19:40:10
@smudgebun:matrix.orgsmudge (she/her) should the nixos wiki cuda page's cache section be corrected? it recommends the cache.nixos-cuda.org cache and doesn't mention others which lead me to assume it was the one for public use. is nix-community.cachix.org (per this announcement) the one that should be primarily recommended on the wiki? as well as information about using flox's nixpkgs repo and its cache? Or, if the nix-community.cachix.org isn't actually allowed to distribute should flox be the primary suggested way to access a cuda cache?

also should the information for cache.nixos-cuda.org be retained, just with a proper disclaimer that it is for internal use only, or how would you like that to be handled?
19:44:22
@smudgebun:matrix.orgsmudge (she/her) should the nixos wiki cuda page's cache section be corrected? it recommends the cache.nixos-cuda.org cache and doesn't mention others which lead me to assume it was the one for public use. is nix-community.cachix.org (per this announcement) the one that should be primarily recommended on the wiki? as well as information about using flox's nixpkgs repo and its cache? Or, if the nix-community.cachix.org isn't technically allowed to distribute should flox be the primary suggested way to access a cuda cache?

also should the information for cache.nixos-cuda.org be retained, just with a proper disclaimer that it is for internal use only, or how would you like that to be handled?
20:10:58
@smudgebun:matrix.orgsmudge (she/her) should the nixos wiki cuda page's cache section be corrected? it recommends the cache.nixos-cuda.org cache and doesn't mention others which lead me to assume it was the one for public use. is nix-community.cachix.org (per this announcement) the one that should be primarily recommended on the wiki? as well as information about using flox's nixpkgs repo and its cache? Or, if the nix-community.cachix.org isn't technically allowed to distribute should flox be the primary suggested way to access a cuda cache?

also should the information for cache.nixos-cuda.org be retained, just with a disclaimer that it is for internal use only?
20:14:14
13 May 2026
@smudgebun:matrix.orgsmudge (she/her) changed their display name from smudge to smudge (she/her).08:48:06
14 May 2026
@weriomat:matrix.orgweriomat joined the room.21:11:04
15 May 2026
@graysontinker:matrix.orgGrayson Tinker changed their display name from graysontinker to Grayson Tinker.03:47:21

Show newer messages


Back to Room ListRoom Version: 9