| 8 Feb 2026 |
 hexa (UTC+1) | it is below the driverLink path | 03:38:20 |
 Robbie Buxton | Yeah on nixos iirc it’s symlinked into /run/opengl-driver/lib if I’m not mistaken | 03:39:39 |
| hexa (UTC+1) | correct | 03:40:14 |
| hexa (UTC+1) | addDriverRunpath.driverLink is the relevant attribute | 03:40:24 |
|  kaya 𖤐 changed their profile picture. | 22:50:15 |
 Gaétan Lepage | After some testing, our current torch version (2.9.0) does build against cuda 13.0, but not cuda 13.1:
/nix/store/42f8i6v4gfkvdimy9aczwqik3scl6dpw-cuda13.1-cuda_cccl-13.1.115/include/cub/device/dispatch/dispatch_radix_sort.cuh(1425): error: no operator "+=" matches these operands
operand types are: at::native::<unnamed>::offset_t += const int64_t
end_offsets_current_it += num_current_segments;
Context: https://github.com/NixOS/nixpkgs/pull/486717 | 23:01:20 |
| Gaétan Lepage | I'll try to ship torch 2.10.0 ASAP, hoping that it is compatible with cuda 13.1 (which should unfortunately * | 23:02:38 |
| Gaétan Lepage | * I'll try to ship torch 2.10.0 ASAP, hoping that it is compatible with cuda 13.1 (which should unfortunately not be the case). | 23:02:57 |
|  @niten:fudo.im left the room. | 23:07:13 |
| 9 Feb 2026 |
 Benjamin Isbarn | I'm not using any overlay for that purpose right now. Good point regarding the global override, will do that ;). So the cudaCapabilities would affect packages like the cudart, cublas etc. I guess i.e. what features it will consider available and thus use? the in theory this should yield better performance for the aforementioned libraries? | 07:03:05 |
 connor (burnt/out) (UTC-8) | Read https://nixos.org/manual/nixpkgs/stable/#cuda -- Jetson isn't built by default and pre-thor uses different binaries so you need to make sure cudaCapabilities is set correctly; you'll get faster builds, smaller closures, and (possibly) better performance if you specify the exact capability | 07:36:40 |
|  SolitudeAlma joined the room. | 07:49:25 |
| Gaétan Lepage | VLLM is now 0.15.1 (latest version) | 11:11:17 |
 cameronraysmith | SomeoneSerge (back on matrix): let me know if the updates to https://github.com/NixOS/nixpkgs/pull/488199 captured what you suggested. No rush: thanks! | 21:03:44 |
| Gaétan Lepage | connor (burnt/out) (UTC-8) SomeoneSerge (back on matrix)
This PR should fix the last failing gpuCheck instance, i.e. python3Packages.triton.gpuCheck: https://github.com/NixOS/nixpkgs/pull/488887
I discovered one of our beloved dlopen instance in triton. We didn't know about it since then... This PR fixes it too. | 23:39:42 |
| Gaétan Lepage | * connor (burnt/out) (UTC-8) SomeoneSerge (back on matrix)
This PR should fix the last failing gpuCheck instance, i.e. python3Packages.triton.gpuCheck: https://github.com/NixOS/nixpkgs/pull/488887
I discovered one of our beloved dlopen instances in triton. We didn't know about it since then... This PR fixes it too. | 23:55:34 |
| 10 Feb 2026 |
| connor (burnt/out) (UTC-8) | Don't think I linked it here, maybe interesting for people with heavy eval jobs: https://github.com/ConnorBaker/nix-optimization | 01:41:36 |
| 11 Feb 2026 |
| connor (burnt/out) (UTC-8) | Gaétan Lepage: there's a merge conflict and I need to rebase, but IIRC something like https://github.com/NixOS/nixpkgs/pull/485208 is necessary to make CUDA 13 the default.
I still need to do the same for the PyCuda PR I have: https://github.com/NixOS/nixpkgs/pull/465047. Apologies that's taking me so long. | 18:56:11 |
| 12 Feb 2026 |
| Gaétan Lepage | Ok thanks! I should get a notification when you'll have rebased. | 07:52:34 |
