| 13 Jan 2026 |
Alexandros Liarokapis | what do I use as inputs and nixpkgs config to get cached ucc with cuda support? | 23:00:25 |
Alexandros Liarokapis | what I did was use https://hydra.nixos-cuda.org/build/63411#tabs-buildinputs for nixpkgs input, then instantiate only with allowUnfree and cudaSupport = true; | 23:01:46 |
Alexandros Liarokapis | nvm my user nix config was overriding my systems's cache | 23:13:26 |
| 14 Jan 2026 |
connor (burnt/out) (UTC-8) | I didn't realize that by making TensorRT 10.14.1 the default I effectively broke aarch64-linux (non Jetson) builds using TensorRT since NVIDIA didn't make a release for that platform for CUDA 12 -- it's CUDA 13 only. So... do I roll back the version of TensorRT on aarch64-linux for CUDA 12.x or do something else? It'd need to get backported to the release branch to unbreak that. 10.13.0 has support for aarch64-linux for CUDA 12, but it was removed in 10.13.2 (there is no 10.13.1 release I can find). | 00:33:23 |
yorik.sar | So they removed support for a whole version+platform in a patch release? That’s not nice. | 04:45:45 |
yorik.sar | I think we’d need 2 versions then - one for aarch64, one for the rest. | 04:46:34 |
hexa (UTC+1) | it might be possible that I can escape the tensorflow curse with ai-edge-litert | 21:30:09 |
hexa (UTC+1) | * it might be possible that I can escape the tensorflow curse with ai-edge-litert, the successor to tflite | 21:30:15 |
| 15 Jan 2026 |
connor (burnt/out) (UTC-8) | Can someone do something to test or verify https://github.com/NixOS/nixpkgs/pull/480232 | 03:07:04 |
hexa (UTC+1) | https://github.com/NixOS/nixpkgs/pull/480103 introduces reintroduces libtensorflow-bin at 2.18.1, if anyone cares for that | 15:15:31 |
hexa (UTC+1) | * https://github.com/NixOS/nixpkgs/pull/480103 introduces reintroduces libtensorflow-bin at 2.18.1, if anyone here cares for that | 15:15:35 |
hexa (UTC+1) |
Note: Publishing standalone libtensorflow packages was dropped in TensorFlow 2.19.0, so the binary version used in this package is the latest one.
| 15:16:07 |
hexa (UTC+1) | ah sure, that's why we package an outdated release 😬 | 15:16:18 |
connor (burnt/out) (UTC-8) | Goddamn stubs and transitive symbol resolution | 17:33:35 |
| * connor (burnt/out) (UTC-8) grumbles | 17:33:39 |
| 16 Jan 2026 |
connor (burnt/out) (UTC-8) | PR to avoid rebuilding Nix itself whenever CUDA support is enabled: https://github.com/NixOS/nixpkgs/pull/480524 | 00:01:13 |
hexa (UTC+1) | nix depends on onetbb? | 00:09:53 |
connor (burnt/out) (UTC-8) | Yes, IIRC at least through blake3 | 02:23:54 |
connor (burnt/out) (UTC-8) | hexa (UTC+1): any additional thoughts on merging https://github.com/NixOS/nixpkgs/pull/480524? | 18:16:50 |
hexa (UTC+1) | none | 18:30:44 |
hexa (UTC+1) | didn't test anything | 18:30:47 |
hexa (UTC+1) | just disables cuda support | 18:30:55 |
hexa (UTC+1) | what could go wrong? | 18:31:03 |
SomeoneSerge (back on matrix) | Wait, globally? | 20:00:50 |
SomeoneSerge (back on matrix) | This probably affects MPI and SLURM, and might be one of those libraries you're actually supposed to link dynamically and not how we do it | 20:26:15 |
connor (burnt/out) (UTC-8) | I don't follow; IIRC hwloc is pulled in by those directly and would use it untouched | 22:48:00 |