!eWOErHSaiddIbsUNsJ:nixos.org

NixOS CUDA

274 Members
CUDA packages maintenance and support in nixpkgs | https://github.com/orgs/NixOS/projects/27/ | https://nixos.org/manual/nixpkgs/unstable/#cuda55 Servers

Load older messages


SenderMessageTime
28 Oct 2025
@connorbaker:matrix.orgconnor (burnt/out) (UTC-8) changed their profile picture.02:19:15
@connorbaker:matrix.orgconnor (burnt/out) (UTC-8) changed their profile picture.02:19:56
@hsngrmpf:matrix.org@hsngrmpf:matrix.org left the room.06:32:54
@connorbaker:matrix.orgconnor (burnt/out) (UTC-8)More progress on the CUDA-legacy repo — started adding more redistributables to older CUDA releases. Found the upper bound for NCCL for CUDA pre-11.7. Building cccl from source, which also means using a newer version, which is fine since they have a good compat story and newer versions allow older CUDA releases to run more recent software. Beginning to get into the pattern of encoding constraints better using brokenAssertions and platformAssertions (see https://github.com/nixos-cuda/cuda-legacy/blob/8fcaae25cb949fa2070f15835e86fa8c0978e712/pkgs/development/cuda-modules/packages/cudnn.nix#L45 and https://github.com/nixos-cuda/cuda-legacy/blob/8fcaae25cb949fa2070f15835e86fa8c0978e712/pkgs/development/cuda-modules/packages/libcublasmp.nix#L52 for examples).07:09:08
@connorbaker:matrix.orgconnor (burnt/out) (UTC-8)It frustrates me so greatly that we don’t have structured warnings or errors and that I had to roll my own with brokenAssertions and platformAssertions and use trace verbose to avoid polluting output with evaluation warnings. And that outside of providing evalHandler when importing Nixpkgs there’s no way to extend such handling. And that the default evalHandler throws a string unrelated to anything it’s processed so adding to the error context doesn’t help.07:12:27
@connorbaker:matrix.orgconnor (burnt/out) (UTC-8)

Two additional TODOs for me today:

  • Make a note in the changelog about why building redistributables requires config.cudaSupport be enabled even though we're just unpacking binary archives (answer: propagates dependency on config.cudaSupport being enabled to consumers of those packages)
  • Make a note about troubleshooting packages being unavailable using eval with --trace-verbose
13:54:51
@connorbaker:matrix.orgconnor (burnt/out) (UTC-8)Could I get a review/approval on https://github.com/NixOS/nixpkgs/pull/45651018:09:36
@glepage:matrix.orgGaétan Lepage🎉19:45:59
@glepage:matrix.orgGaétan Lepageclipboard.png
Download clipboard.png
19:46:00
@glepage:matrix.orgGaétan Lepageclipboard.png
Download clipboard.png
20:26:41
@glepage:matrix.orgGaétan LepagePeak performance packaging by the pytorch folks20:26:56
@sporeray:matrix.orgRobbie BuxtonOmg20:27:54
@glepage:matrix.orgGaétan Lepageclipboard.png
Download clipboard.png
20:28:39
29 Oct 2025
@connorbaker:matrix.orgconnor (burnt/out) (UTC-8)Thanks, I hate it01:08:50
@connorbaker:matrix.orgconnor (burnt/out) (UTC-8)

1 PR merged fixing nsight_systems GUI, 3 PRs merged with package expression tidying and support for more releases (from CUDA legacy), and 1 PR ready to go for the changelog: https://github.com/NixOS/nixpkgs/pull/456510

There are still a bunch of redist packages to introduce in CUDA legacy and find requirements for :/

Also still need to make a PR to drop all the packages with “withCuda” variants…

And then there’s the setup hook work 🫩

04:40:06
@connorbaker:matrix.orgconnor (burnt/out) (UTC-8)Is that good?04:40:42
@glepage:matrix.orgGaétan LepageYes and no, the "+125" is mostly the new stuff introduced by the CUDA 13 PR.08:48:19
@connorbaker:matrix.orgconnor (burnt/out) (UTC-8)This may be worth considering: https://github.com/garnix-io/yensid14:56:27
@ss:someonex.netSomeoneSerge (back on matrix)Neat, literally haproxy16:00:07
@glepage:matrix.orgGaétan Lepage

FYI, the update to the latest torch 2.9.0 was just merged!

Changelog: https://github.com/pytorch/pytorch/releases/tag/v2.9.0

We also just added torchcodec a media decoder/encoder working along with PyTorch

23:19:33
30 Oct 2025
@glepage:matrix.orgGaétan Lepage connor (he/him) (UTC-7) any idea what, in the CUDA 13 PR, could be responsible for gcc-wrapper leaking into the firefox derivation? 21:42:34
@glepage:matrix.orgGaétan Lepage
❮ nix-build --arg config '{ allowUnfree = true; cudaSupport = true; }' -A firefox
this derivation will be built:
  /nix/store/7s1dyndj7r33wck6dywjf6kqxah7ngm2-firefox-144.0.2.drv
building '/nix/store/7s1dyndj7r33wck6dywjf6kqxah7ngm2-firefox-144.0.2.drv'...
structuredAttrs is enabled
error: output '/nix/store/b3cyrj28jf03mrf15jmy9whhfrgy2bxw-firefox-144.0.2' is not allowed to refer to the following paths:
         /nix/store/x8mydcgbry214s802nzvy7fdljx404ym-gcc-wrapper-14.3.0
21:42:54
@connorbaker:matrix.orgconnor (burnt/out) (UTC-8)Check onnxruntime22:04:03
@glepage:matrix.orgGaétan LepageBefore:22:11:06
@glepage:matrix.orgGaétan Lepageclipboard.png
Download clipboard.png
22:11:08
@glepage:matrix.orgGaétan LepageAfter:22:11:19
@glepage:matrix.orgGaétan Lepageclipboard.png
Download clipboard.png
22:11:21
@glepage:matrix.orgGaétan Lepage Probably nccl's fault... 22:18:18
31 Oct 2025
@connorbaker:matrix.orgconnor (burnt/out) (UTC-8) changed their profile picture.03:16:13
@connorbaker:matrix.orgconnor (burnt/out) (UTC-8) changed their display name from connor (he/him) (UTC-7) to connor (burnt/out) (UTC-7).03:16:45

Show newer messages


Back to Room ListRoom Version: 9