| 22 Dec 2024 |
 SomeoneSerge (back on matrix) | 👍️ cuda and rocm in pytorch are mutually exclusive, so I guess there must be a separate wheel | 00:19:20 |
 matthewcroughan | https://hydra.nixos.org/job/nixpkgs/trunk/zluda.x86_64-linux | 10:51:11 |
| matthewcroughan | when is the last time this even built | 10:51:18 |
| matthewcroughan | it seems like literally never? | 10:51:30 |
| matthewcroughan | how was this merged if it never built? | 10:51:34 |
| SomeoneSerge (back on matrix) | Back when rocmPackages weren't broken | 11:28:26 |
| SomeoneSerge (back on matrix) | * Back when rocmPackages weren't broken exceeding the hydra quotas | 11:28:39 |
| matthewcroughan | SomeoneSerge (utc+3): I would love to try out zluda, but do not want to wait over 7 hours for the build, or whatever it takes | 12:21:49 |
| matthewcroughan | got a cache? | 12:21:52 |
| SomeoneSerge (back on matrix) | No, sorry. I haven't got any amd hardware at home unfortunately | 12:37:10 |
| matthewcroughan | Yeah I think there's a lack of investment in the AMD community. | 12:37:30 |
| matthewcroughan | In a fairer world, we would be given the hardware to support this, but we don't have such a thing. And given that nobody is willing to give, then they won't get. | 12:37:49 |
| matthewcroughan | The state of things, I think reflects a lack of support from the people who need to give it. | 12:38:27 |
| matthewcroughan | And I don't see many people complaining either, so maybe there aren't very many users anyway? | 12:38:54 |
| SomeoneSerge (back on matrix) | I still think the clusters are the main user, and these people currently aren't sold on "this kind of reproducibility" | 13:22:51 |
|  @stablejoy:matrix.org left the room. | 13:24:59 |
| matthewcroughan | What I mean is that I could sit here and build stuff, and with some hard drives given to me host a cache for almost free on a gigabit connection | 13:30:52 |
| matthewcroughan | But there's no incentive for me to do so, because I would starve trying | 13:31:01 |
| SomeoneSerge (back on matrix) | Nooooo you have to build a profitable business that justifies building stuff, embrace the grind | 14:15:23 |
| matthewcroughan | I think what I'm saying is that there seems to be exactly no way to justify building for AMD, and that this goes beyond NixOS and upstream | 14:20:45 |
|  allrealmsoflife joined the room. | 15:55:31 |
| 23 Dec 2024 |
 hexa | is anyone looking into the triton-llvm test issues on stagin-next? | 13:40:22 |
| hexa | * is anyone looking into the triton-llvm test issues on staging-next? | 13:40:25 |
| hexa | https://hydra.nixos.org/build/281892957 | 13:40:30 |
| 24 Dec 2024 |
| matthewcroughan | Hah, this old good one | 11:39:47 |
| matthewcroughan | SomeoneSerge (utc+3): you might be interested in https://github.com/huggingface/optimum-quanto/issues/361 | 11:40:09 |
| matthewcroughan | Because optimum-quanto depends on a string that only exists in torch-bin
And when using torch-bin in an overlay, this happens
error: attribute 'cudaPackages' missing
at /nix/store/6zgbbqlr7nnfxpzkyj7fsl4fpg89jbw0-source/pkgs/development/python-modules/bitsandbytes/default.nix:16:19:
15|
16| inherit (torch) cudaPackages cudaSupport;
| ^
17| inherit (cudaPackages) cudaVersion;
| 11:40:27 |
| matthewcroughan | And when I tell this specific package to use torch-bin, it turns out torch-bin doesn't seem to compile because of python3.12-triton> chmod: cannot access '/nix/store/kwahd53lf8hqylcpiyv15l157iayn1gc-python3.12-triton-3.1.0/lib/python3.12/site-packages/triton/third_party/cuda/bin/ptxas': No such file or directory | 11:41:32 |
| SomeoneSerge (back on matrix) | Hmmm there used to be a postPatch in triton-bin symlinking our ptxas | 13:54:40 |
| SomeoneSerge (back on matrix) | * Hmmm there used to be a postPatch in triton-bin (and in triton) symlinking our ptxas | 13:54:52 |
