| 30 Jan 2025 |
 ruro | * I occasionally have some free time and I would like to spend it on improving the state of CUDA in nixpkgs/NixOS. Do you have any suggestions for which issues I should start with? Alternatively, I could just start going down the list of eval/build failures on the nix-community CUDA builder... | 16:24:46 |
| ruro | * I occasionally have some free time and I would like to spend it on improving the state of CUDA in nixpkgs/NixOS. Do you have any suggestions for which issues I should start with?
I am not a complete newbie, but I haven't contributed to nixpkgs all that much (above PR is my third). I guess, I could just start going down the list of eval/build failures on the nix-community CUDA builder... but that might end up just being treadmill work. | 16:28:22 |
| ruro | * I occasionally have some free time and I would like to spend it on improving the state of CUDA in nixpkgs/NixOS. Do you have any suggestions for which issues I should start with?
I am not a complete newbie, but I haven't contributed to nixpkgs all that much (the above-mentioned PR is my third). I guess, I could just start going down the list of eval/build failures on the nix-community CUDA builder... but that might end up just being treadmill work. | 16:47:05 |
 connor (he/him) | In reply to @glepage:matrix.org
Looks like we need to wait for https://github.com/NixOS/nixpkgs/pull/306172 I meant to close that since I’ll make a new PR for the CUDA-packages work | 16:47:34 |
 Gaétan Lepage | Ok Connor. Do we have a short-term alternative to get this library ? | 16:48:28 |
| connor (he/him) | Depends on what you mean by short term :( | 16:53:34 |
| connor (he/him) | I should have everything landed by 25.05 but I suppose we’ll need it prior to that | 16:54:03 |
| connor (he/him) | I guess I can start trying to land things, but it’ll cause some breakages and I don’t have docs written yet | 16:54:43 |
| Gaétan Lepage | I'm asking for pytorch (https://github.com/NixOS/nixpkgs/pull/377785). There is no emergency and we can surely wait before updating it. | 16:56:39 |
| connor (he/him) | Ugh didn’t they also remove support for CUDA 12.1 | 16:58:19 |
| connor (he/him) | Also I think they support newer architectures now (maybe Blackwell?) | 16:58:36 |
| Gaétan Lepage | In reply to @connorbaker:matrix.org
Ugh didn’t they also remove support for CUDA 12.1 From the CI at least: https://github.com/pytorch/pytorch/pull/141271, https://github.com/pytorch/pytorch/pull/142177 | 16:59:51 |
| connor (he/him) | My bad I mixed up the CI removal and https://github.com/NVIDIA/TensorRT-Model-Optimizer/releases/tag/0.23.0 removing support for CUDA 11 | 17:02:04 |
| 31 Jan 2025 |
 Kevin Mittman (UTC-8) | In reply to @glepage:matrix.org
Which package provides libcusparseLt.so.0 ? It's separate https://developer.download.nvidia.com/compute/cusparselt/redist/ | 02:13:02 |
| connor (he/him) | I am so tired
But now have setup hooks which can catch common issues like the order of different CUDA directories in a run path
Or fail a build if NVCC’s host compiler leaks out (which can/will cause glibc/glibcxx symbol issues)
Even beyond that
I implemented utility functions for arrays and associative arrays in bash because I got tired of repeating myself in different hooks
And then when I got tired of repeating myself in tests for those functions and hooks, I made a utility derivation to make testing for expected arrays and associative arrays easier | 06:55:57 |
| connor (he/him) | It’s still a mess but it’s on this branch if anyone is curious https://github.com/ConnorBaker/cuda-packages/compare/main...fix/runpath-order-matters-and-cuda-compat-gets-clobbered | 06:56:57 |
 SomeoneSerge (back on matrix) | Let's schedule a call to discuss how to go forward with stdenv support, setup-hooks, wrappers, config.cudaSupport, localSystem/pkgsCross, clang support, and out-of-tree override-ability of manifests and toolkit components https://crab.fit/cudapackages-ng-781527
CC connor (he/him) (UTC-7), sielicki, Samuel Ainsworth, and anyone interested
| 10:49:09 |
|  Jonas Chevalier changed their display name from Jonas Chevalier to Jonas Chevalier (FOSDEM). | 19:11:42 |
| SomeoneSerge (back on matrix) changed their display name from SomeoneSerge to SomeoneSerge (Bruxelles). | 19:35:00 |
| 1 Feb 2025 |
|  matthewcroughan changed their display name from matthewcroughan (already in Brussels) to matthewcroughan (FOSDEM). | 09:41:01 |
| 2 Feb 2025 |
|  pbsds changed their display name from pbsds to pbsds (FOSDEM). | 16:04:38 |
|  osbm joined the room. | 18:23:02 |
| 3 Feb 2025 |
| Jonas Chevalier changed their display name from Jonas Chevalier (FOSDEM) to Jonas Chevalier. | 08:23:12 |
| matthewcroughan changed their display name from matthewcroughan (FOSDEM) to matthewcroughan. | 09:11:41 |
| SomeoneSerge (back on matrix) changed their display name from SomeoneSerge (Bruxelles) to SomeoneSerge (Gand St. Pieters). | 13:40:41 |
| ruro | connor (he/him) (UTC-7): SomeoneSerge (Gand St. Pieters) sorry to keep annoying you guys, but could you respond to the above question? Alternatively, "we are too busy right now, you'll have to figure it out on your own" is also an acceptable answer))) | 14:37:45 |
| SomeoneSerge (back on matrix) | Sorry, I forgot to reply. I'll write before tomorrow | 14:41:33 |
| ruro | ❤️ | 14:42:10 |
| pbsds changed their display name from pbsds (FOSDEM) to pbsds. | 16:25:49 |
| SomeoneSerge (back on matrix) | Starting with the last question: great to hear! As one tool to help with discovery, we have a task board at https://github.com/orgs/NixOS/projects/27/views/1. We haven't been properly maintaining it for the last year, I see many invalidated/outdated items there, but some of the roadmap is still relevant, and the "New" column is automatically populated with all issues and PRs tagged "cuda".
If you're willing to do chores, fixing issues like "nvidia's bash wrapper for nsys-ui assumes things are installed into weird locations and is completely broken" and "a package has changed the way they hardcode /usr/lib or dlopen stuff and now fails to find libcuda.so again", those would be very useful, relatively straightforward, but involve an amount of debugging and suffering and usually get ignored for a long time because it's just demotivating.
If you're interested in architectural issues, then note the message about the upcoming meeting and the proposed subjects, check out the "Roadmap" column, and Connor's out-of-tree cuda-packages
| 22:27:33 |
