| 18 Mar 2025 |
 stick | * if you are feeling adventurous, you might as well update pkgs/development/cuda-modules/tensorrt/releases.nix to contain the latest 10.9.x.y release (both for cuda 11.x and 12.x) | 20:12:16 |
 Michal Koutenský | both report '10.3.0' for me | 20:13:44 |
| Michal Koutenský | sure, i can do that tomorrow | 20:14:26 |
| stick | in that case it makes sense to update version on line 16, not the unpack phase | 20:15:01 |
| stick | so the nix package has the same version as reported by wheel | 20:15:18 |
| Michal Koutenský | yeah that makes sense | 20:15:26 |
| stick | SomeoneSerge (UTC+U[-12,12]): I finally managed to build onnxruntime with CUDA 12.8 - see https://github.com/NixOS/nixpkgs/pull/390885
can you run the CUDA version bump test suite and let me know if we can merge the PR?
| 22:20:53 |
| stick | * SomeoneSerge (UTC+U[-12,12]): I finally managed to build onnxruntime (afaik the only blocker) with CUDA 12.8 - see https://github.com/NixOS/nixpkgs/pull/390885
can you run the CUDA version bump test suite and let me know if we can merge the PR?
| 22:21:20 |
| stick | I am building magma/torch/vllm as we speak - but IIRC this already went OK when i tried it few weeks ago | 22:21:59 |
| stick | the onnxruntime fix was to turn off LTO - it went bonkers (into infinite loop) when trying to do the final link - I looked into gentoo ebuilds and they also turn off LTO when linking onnxruntime with cuda | 22:24:01 |
 SomeoneSerge (back on matrix) | stick: just a heads up: I've been using the university workstation for nixpkgs-review so far, but I'm no longer employed by the uni and am migrating between infrastructures; going to take time 🤷 | 22:24:51 |
| stick | so the version bump test is only running nixpkgs-review on the PR? | 22:25:57 |
| SomeoneSerge (back on matrix) | There's, nixpgks-review with cudaSupport=true, there's passthru gpuChecks, there's samuela's and Connor's out-of-tree test-suites | 22:27:50 |
| SomeoneSerge (back on matrix) | Nixpkgs-review is a bit of cargo culting but it gives an idea of the size of the fallout | 22:29:14 |
| stick | yeah, i use it often locally - but there are many failing packages also on master unfortunately | 22:30:00 |
| stick | * yeah, i use it often locally - but there are many not-so-important failing packages also on master unfortunately | 22:30:20 |
| SomeoneSerge (back on matrix) | Yes, it's definitely too much compute hours spent on just getting a statistic (# failures) | 22:38:34 |
| 21 Mar 2025 |
|  Domen Kožar changed their profile picture. | 11:40:19 |
| 23 Mar 2025 |
|  connor (he/him) changed their display name from connor (he/him) (UTC-8) to connor (he/him) (UTC-7). | 18:28:18 |
| 24 Mar 2025 |
| connor (he/him) | How did I not know about nvCOMP?! https://docs.nvidia.com/cuda/nvcomp/index.html | 14:37:00 |
| SomeoneSerge (back on matrix) | Did you ever need it? | 14:37:53 |
 Moritz Sanft | Are you sure that removing the OpenSSL1.1 compatibility libraries from the NVIDIA driver package is correct?
see: https://github.com/NixOS/nixpkgs/blob/a13cc0bc42a89a025f545c4a17f25e6300710f39/pkgs/os-specific/linux/nvidia-x11/builder.sh#L81-L82
I ran into PyTorch 2.6 trying to open that library today, hence the question. | 15:41:42 |
| Moritz Sanft | * Are we sure that removing the OpenSSL1.1 compatibility libraries from the NVIDIA driver package is correct?
see: https://github.com/NixOS/nixpkgs/blob/a13cc0bc42a89a025f545c4a17f25e6300710f39/pkgs/os-specific/linux/nvidia-x11/builder.sh#L81-L82
I ran into PyTorch 2.6 trying to open that library today, hence the question. | 15:41:49 |
| SomeoneSerge (back on matrix) | Hm. The user would have to manually enable patchelfing the paths to openssl 1.1 anyway, I'd say we shouldn't include them y default because then everything goes behind permittedInsecurePackages or what's the name | 15:43:20 |
| SomeoneSerge (back on matrix) |
I ran into PyTorch 2.6 trying to open that library today, hence the question.
Can you also file a report with the logs on github?
| 15:43:58 |
| Moritz Sanft | Sure! | 15:47:58 |
 Gaétan Lepage | Hi,
I'm looking at mistral-rs and trying to make autoAddDriverRunPath and autoPatchelfHook work, but I get this at runtime:
Unable to dynamically load the "cuda" shared library - searched for library names: ["cuda", "nvcuda"]. Ensure that `LD_LIBRARY_PATH` has the correct path to the installed library. If the shared library is present on the system under a different name than one of those listed above, please open a GitHub issue.
| 16:30:22 |
| SomeoneSerge (back on matrix) | They probably CDLL on the python side? | 17:26:26 |
| Gaétan Lepage | It's rust | 17:43:35 |
| Gaétan Lepage | But yeah, they surely DLopen | 17:43:40 |
