!eWOErHSaiddIbsUNsJ:nixos.org

NixOS CUDA

290 Members
CUDA packages maintenance and support in nixpkgs | https://github.com/orgs/NixOS/projects/27/ | https://nixos.org/manual/nixpkgs/unstable/#cuda58 Servers

Load older messages


SenderMessageTime
10 Feb 2025
@stick:matrix.orgsticksounds good 15:05:28
@ruroruro:matrix.orgruro

Btw, it seems that there is a mismatch in CUDA vs GCC version compatibility after all. In particular, for me

cudaPackages_11_2.backendStdenv.cc.version == "12.4.0"

and with that compiler, a couple of the samples in cuda-samples are failing to compile due to errors inside cuda_cccl headers. Here is a forum post discussing this exact error.

I wasn't able to find precise compatibility matrices for CUDA 12.1, but according to this stackoverflow answer, CUDA 12.1 supports GCC versions up to 12.2 which is lower than the current 12.4.0.

Minor versions don't work in pkgs/development/cuda-modules/nvcc-compatibilities.nix, but setting gccMaxMajorVersion = "11" for "12.1" I was able to successfully build the problematic samples that fail with GCC 12.4.0. So it looks like CUDA 12.1 doesn't fully support GCC 12.4 after all.

Thoughts?

15:53:25
@ss:someonex.netSomeoneSerge (back on matrix)Hm. The logic for choosing gcc initially used gccMaxMajorVersion from some other file, nvcc-compatibilities was added later15:54:53
@ss:someonex.netSomeoneSerge (back on matrix)I hope to rewrite the eval bit in the coming month15:55:19
@ruroruro:matrix.orgruro *

Btw, it seems that there is a mismatch in CUDA vs GCC version compatibility after all. In particular, for me

cudaPackages_11_2.backendStdenv.cc.version == "12.4.0"

and with that compiler, a couple of the samples in cuda-samples are failing to compile due to errors inside cuda_cccl headers. Here is a forum post discussing this exact error.

I wasn't able to find precise compatibility matrices for CUDA 12.1, but according to this stackoverflow answer, CUDA 12.1 supports GCC versions up to 12.2 which is lower than the current 12.4.0.

Minor versions don't work in pkgs/development/cuda-modules/nvcc-compatibilities.nix, but setting gccMaxMajorVersion = "11" for CUDA "12.1" I was able to successfully build the problematic samples that fail with GCC 12.4.0. So it looks like CUDA 12.1 doesn't fully support GCC 12.4 after all.

Thoughts?

15:57:11
@stick:matrix.orgsticki was able to build most of the release-cuda packages on aarch64-linux without any build issues - so this is awesome!17:29:24
@stick:matrix.orgsticki pushed the results into https://app.cachix.org/cache/cuda-arm in case anyone would like to play with that17:30:29
@stick:matrix.orgstickbest it would be to have aarch64-linux enabled for nixpkgs:cuda on nix-community ofc17:30:54
@mcwitt:matrix.orgmcwitt

Hi all, I've been wanting to use compute-sanitizer from cudaPackages.cuda_sanitizer_api but noticed it's currently broken (reported here: https://github.com/NixOS/nixpkgs/issues/349426).

In the issue SomeoneSerge (UTC+U[-12,12]) gave a useful hint for how to get started fixing this. I've implemented the suggestion in https://github.com/NixOS/nixpkgs/pull/379695.

This does seem like a useful starting point: the compute-sanitizer binary is working and I've been successfully using it at work. Is anyone here available to review?

23:01:35
12 Feb 2025
@connorbaker:matrix.orgconnor (burnt/out) (UTC-8)Just merged it, thank you!15:32:37
@connorbaker:matrix.orgconnor (burnt/out) (UTC-8)All, I’m excited for our meeting tomorrow! Do we have a document with an agenda or minimum set of items to cover?15:33:17
@ss:someonex.netSomeoneSerge (back on matrix) Me too! Just started one https://pad.lassul.us/YGyymxE9Qqy9iFVt7A2VnA#, everyone feel free to edit 20:09:57
13 Feb 2025
@connorbaker:matrix.orgconnor (burnt/out) (UTC-8) changed their display name from connor (he/him) (UTC-7) to connor (he/him) (UTC-8).06:59:16
@ss:someonex.netSomeoneSerge (back on matrix)Matrix calls did work last time didn't it?07:51:04
@ss:someonex.netSomeoneSerge (back on matrix) * Matrix call did work last time didn't it?07:51:09
@ss:someonex.netSomeoneSerge (back on matrix)Still haven't figured out what broke in jitsi07:51:29
@ss:someonex.netSomeoneSerge (back on matrix) This is in 50' from now 13:11:36
@ss:someonex.netSomeoneSerge (back on matrix) Let's try this I suppose 13:55:54
Jitsi widget added by @ss:someonex.netSomeoneSerge (back on matrix)13:56:07
@ss:someonex.netSomeoneSerge (back on matrix)(I wonder if this sends room-wide notifications 🤔)13:56:28
@connorbaker:matrix.orgconnor (burnt/out) (UTC-8)I didn't see a notification, but I did see it pop up in the chat13:56:49
@palasso:matrix.org@palasso:matrix.orgIt does. I got a notification.14:02:01
@srhb:matrix.orgsrhbYup :D 14:02:10
@connorbaker:matrix.orgconnor (burnt/out) (UTC-8)

Gaétan Lepage: I've got the manifest for cusparseLT here: https://github.com/ConnorBaker/cuda-packages/blob/main/modules/redists/cusparselt/manifests/0.6.3.json

I think with that you should be able to construct a Nix expression which manually calls redist-builder (or whatever I called it upstream) with the proper arguments

15:24:04
@ss:someonex.netSomeoneSerge (back on matrix)Meeting notes: https://pad.lassul.us/YGyymxE9Qqy9iFVt7A2VnA?both#Conclusion. Some intermediate conversations missing right now, but are recorded by Connor; hopefully he can fill in the blanks when he's free15:30:22
@connorbaker:matrix.orgconnor (burnt/out) (UTC-8)Just pasting the last of them now15:30:46
@ss:someonex.netSomeoneSerge (back on matrix)

Regarding scheduling the future meetings,

  • we should probably aim to meet in 2-4 weeks to follow up on the patchelf exception and for a report on the ephemeral builders situation;
  • we can probably first bring up the alignment questions with nix-community just in their chat, without video because async is faster;
  • additionally, I think I should have hours this and next week to sort the backlog as mentioned in the notes; I think it'd still be useful, for onboarding new people, to do that with the audio and the screenshare, but it's not worth synchronizing people's schedules for this; maybe it'll be just a pop-in format?
15:38:04
@ss:someonex.netSomeoneSerge (back on matrix)(jaja, maybe we do this in Gaetan's twitch?)15:38:26
@glepage:matrix.orgGaétan LepageSure haha15:46:34
14 Feb 2025
@connorbaker:matrix.orgconnor (burnt/out) (UTC-8)As of a few days ago Onnxruntime requires CUDA separable compilation… so I guess I gotta fix that now 🙃01:50:24

Show newer messages


Back to Room ListRoom Version: 9