| 8 Jun 2024 |
 SomeoneSerge (back on matrix) | Also would it be such a good idea to restart another drama now of all times 😅 | 11:13:12 |
 hexa (UTC+1) | I don't think it is a drama | 11:19:35 |
| hexa (UTC+1) | it is a choice about how useful nixpkgs will be for gpu compute | 11:19:55 |
| hexa (UTC+1) | the current cache hit ratio isn't too great | 11:20:59 |
| hexa (UTC+1) | the drama here is, that I have build faiss right now | 11:22:20 |
| hexa (UTC+1) | In reply to @hexa:lossy.network
possible, varies between minutes and hours 😕 just built the cuda-enabled version in 8m55s | 12:12:47 |
| hexa (UTC+1) | the test probably has a pathological issue, when it is resource straved | 12:13:11 |
| hexa (UTC+1) | * the test probably has a pathological issue, when it is resource starved | 12:13:13 |
| hexa (UTC+1) | * the test suiteprobably has a pathological issue, when it is resource starved | 12:13:20 |
| SomeoneSerge (back on matrix) | Yea -j1 helps with a lot of them... | 12:13:35 |
| SomeoneSerge (back on matrix) | In reply to @hexa:lossy.network
the current cache hit ratio isn't too great I agree, this needs to be done, I want this to be done... | 12:14:31 |
| hexa (UTC+1) | * the drama here is, that I have to build faiss right now | 12:58:30 |
| hexa (UTC+1) | hm | 14:16:26 |
| hexa (UTC+1) | build faiss twice now, the build terminates and there is no output | 14:16:38 |
| hexa (UTC+1) | 😕 | 14:16:45 |
| hexa (UTC+1) | what is this sorcery | 14:16:51 |
| hexa (UTC+1) | ok, third try is the charm | 16:36:45 |
| hexa (UTC+1) | pytorch-metric-learning crashes the test when faiss is built with cudaSupport 🤔 | 16:50:49 |
| hexa (UTC+1) | can someone confirm that? | 17:06:40 |
| hexa (UTC+1) | blurgh, hours of build time down the drain, because whisper-ctranslate2's speaker diarization feature crashes in openblas | 18:17:02 |
| hexa (UTC+1) | make sure your libraries build on python3.12, we're switching to 3.12 as the primary python version in this cycle | 18:19:38 |
| hexa (UTC+1) | like faiss https://github.com/NixOS/nixpkgs/pull/313192 | 18:22:23 |
 Gaétan Lepage | What do you think about marking torchWithRocm as broken ?
https://github.com/NixOS/nixpkgs/pull/318338 | 21:14:07 |
| Gaétan Lepage | It would sanitize the nixpkgs-review output for the torch update | 21:14:57 |
| SomeoneSerge (back on matrix) | In reply to @glepage:matrix.org
What do you think about marking torchWithRocm as broken ?
https://github.com/NixOS/nixpkgs/pull/318338 Didn't read but imo yes,this just adds knowledge, and they.should unmark it once it's ffixed | 21:30:28 |
| Gaétan Lepage | In reply to @ss:someonex.net
Didn't read but imo yes,this just adds knowledge, and they.should unmark it once it's ffixed Yes, I agree. It makes things clearer | 21:36:52 |
| hexa (UTC+1) | there were two rocm maintainers some time ago | 22:34:09 |
| hexa (UTC+1) | ideally they would chime in and clarify their commitment | 22:34:18 |
 connor (burnt/out) (UTC-8) | In reply to @hexa:lossy.network
make sure your libraries build on python3.12, we're switching to 3.12 as the primary python version in this cycle Oof does PyTorch Dynamo even work with 3.12 yet? I haven’t checked since 2.1 I think | 23:05:58 |
| 9 Jun 2024 |
| SomeoneSerge (back on matrix) | Hmm there's a PR starting with python3Packages.torch: but Ofborg didn't try building it, just shows 20 green (eval&c) checks | 00:36:09 |
