| 16 Apr 2026 |
 nim65s | * but you build with how many cores ? | 13:37:31 |
 hexa | b) without excessive cores the most ram usage is during linking | 13:37:38 |
| hexa | I'd say firefox takes at most 40 GB to link | 13:37:55 |
 Robbie Buxton | I was using nixbuild.net and it kept sizing it up and giving it more cores, and yeah the solution was reducing cores obviously | 13:38:02 |
| hexa | that's not friendly at all, but does fit on my desktop still | 13:38:12 |
| hexa | but that means your core/ram balance was way off | 13:38:28 |
| Robbie Buxton | In reply to @hexa:lossy.network
but that means your core/ram balance was way off How so? | 13:39:20 |
| hexa | you have a base need of RAM per core | 13:39:51 |
| hexa | just to run the build on it | 13:39:58 |
| hexa | and if you oom the total cores probably needed more ram than what was provided | 13:40:17 |
| hexa | and the solution would have been to get fewer cores with more ram instead | 13:40:30 |
| Robbie Buxton | Idk I canāt get more than 1.6 TB of ram on AWS š | 13:40:42 |
| hexa | now I don't know anything about how nixbuild.net does scaling | 13:40:56 |
| Robbie Buxton | The large instances have interesting ratios | 13:40:55 |
| hexa | yeah, but if you get 10k cores at the same time, then what? | 13:41:09 |
| Robbie Buxton | Iirc itās not that smart for sizing (the service itself is fab tho) | 13:41:32 |
| hexa | that's 160 MB per core if I didn't miscalculate | 13:41:38 |
| Robbie Buxton | In reply to @hexa:lossy.network
yeah, but if you get 10k cores at the same time, then what? On a single machine I can dream | 13:41:47 |
| hexa | right | 13:41:55 |
 GaƩtan Lepage | Yes, our cuda-packages jobset is broken because python3Packages.deep-ep emits a warning at eval time. | 13:56:06 |
| GaƩtan Lepage | If someone may review https://github.com/NixOS/nixpkgs/pull/510375, we could start building again | 13:56:23 |
| GaƩtan Lepage | Sorry everyone for the disruption | 13:56:35 |
 SomeoneSerge (matrix works sometimes) | A reminder that the cache is not an officially maintained cuda team thing at the moment, and only exposed as an accidental dev artifact. Reminder that "stable channels" are not an official cuda team thing. Only hydra itself is. For the time being. | 14:00:03 |
| 17 Apr 2026 |
| GaƩtan Lepage | SomeoneSerge (matrix works sometimes) do you think we can get this in today?
https://github.com/NixOS/nixpkgs/pull/498612#issuecomment-4266520526
It's be nice to have it as part as 26.05 and the window closes by tomorrow | 09:31:18 |
| SomeoneSerge (matrix works sometimes) | These changes are backportable | 12:59:28 |
| SomeoneSerge (matrix works sometimes) | * | 12:59:47 |
| GaƩtan Lepage | Sure, but it'd be easier to merge them now (if they're ready ofc)c | 13:11:32 |
| GaƩtan Lepage | * Sure, but it'd be easier to merge them now (if they're ready ofc) | 13:11:34 |
| 18 Apr 2026 |
 ccicnce113424 | Actually, the window doesn't close until 2026-04-27 07:00 UTC, so we've still got a few days, not "tomorrow". That said, I think this PR is ready to merge right now. | 09:03:52 |
| 20 Apr 2026 |
 Edward Tjƶrnhammar | You risk breaking HPC deployments with these changes | 09:05:45 |
