| 5 Nov 2025 |
 apyh | couple neat talks but nothing revolutionary | 03:33:57 |
 Robbie Buxton | I will say tho @apyh for packaging I often disable high capacities until after I’ve got it working to make it quicker to interate. | 03:38:39 |
| Robbie Buxton | I’ve yet to hit a capability specific build issue touch wood | 03:38:54 |
 connor (he/him) | Merged https://github.com/NixOS/nixpkgs/pull/458835
hexa (UTC+1) that should fix the Firefox breakage due to CUDA leaking into the default | 15:33:50 |
 hexa | thannks | 15:36:24 |
| hexa | * thanks | 15:36:26 |
 Gaétan Lepage | I have the pleasure to tell you that our infra has churned through the new cuda-packages-v2 jobset for the first time!
https://hydra.nixos-cuda.org/jobset/cuda/cuda-packages-v2 | 23:31:59 |
| 6 Nov 2025 |
| connor (he/him) | Composition of overlays is associative why does lib use foldr 🤨 | 04:27:40 |
| connor (he/him) | * | 04:28:25 |
 Daniel Fahey | vLLM devs want to use triton_kernels (a Python library that's in the Triton monorepo). What would be our advice to them re: "vendoring" (other than "don't")? https://github.com/vllm-project/vllm/issues/27672#issuecomment-3497537863 | 15:08:53 |
| 7 Nov 2025 |
| connor (he/him) | so glad I spent several weeks writing a bunch of bash array utility functions months ago | 09:40:03 |
| Daniel Fahey | have you heard of python? | 09:40:40 |
| connor (he/him) | was able to re-use stuff from https://github.com/connorbaker/cuda-packages and https://github.com/ConnorBaker/nixpkgs/tree/feat/arrayUtilities-remaining to make https://github.com/NixOS/nixpkgs/pull/459416 | 09:41:10 |
| Daniel Fahey | 😜 | 09:41:28 |
| connor (he/him) | using python meaningfully for setup hooks is something I want to push further with the stdenv team but... there's a great deal of different ideas in terms of what that would look like | 09:43:05 |
| connor (he/him) | some want to move away from shell (or to a different shell) for setup hooks, others want to enable more radical changes like the ability to use python for setup hooks or even to replace some or all of setup.sh | 09:44:29 |
| Daniel Fahey | I wrote this recently https://github.com/NixOS/nixpkgs/pull/448828/files#diff-e82c6e9710143cae7452bab94f76c5adedbfc093ca951aa0a0f634d661ef5833 | 09:44:46 |
| connor (he/him) | there's been a big effort to get python3Minimal usable within the bootstrap iirc | 09:44:52 |
| connor (he/him) | okay I need to sleep | 09:46:52 |
| Daniel Fahey | gn gg | 09:47:35 |
| connor (he/him) | SomeoneSerge (back on matrix) Gaétan Lepage this is kind of what I mentioned in the weekly meeting about needing to remove the stubs from the runpath; it doesn't do everything but it's a start. For example, this should unbreak packages which link against stubs and don't use autoAddDriverRunpath --in that case the runpath wouldn't be prefixed by /run/opengl-driver/lib so the stubs would be discovered and loaded and runtime would fail. | 09:48:48 |
| Gaétan Lepage | Ok neat! | 09:53:37 |
| Daniel Fahey | Really, really high quality bash, fan of Wooledge I presume? And the integration with the Nix for the testing 👨🍳 😘, thanks for exposing me to this. I did a lot of bash in my old job inside Makefiles, looking at this has made me remember how much I love bash. Sorry for my tongue in cheek Python remark. | 11:01:25 |
 SomeoneSerge (back on matrix) | This sounds like bootstrap hell? | 13:22:29 |
| SomeoneSerge (back on matrix) | Gonna read, but what I wanted to ask when you were leaving was that we had already introduced some complex bash logic for ensuring the order of hooks, which was meant to ensure that adddriversrunpath and addcudacompat take priority over stubs, which must have clearly regressed if you found cleaning runpaths necessary, so why not fix that instead?
Licensing issues aside, I'd rather have stubs with customized messages linked...
| 13:28:03 |
| SomeoneSerge (back on matrix) | * | 13:34:14 |
| connor (he/him) | The logic I was thinking of that did that was never merged IIRC
I had implemented it in my CUDA-packages repo but it required many more of the array utility functions I wrote to do hook de-ordering, filtering, removal of duplicates, etc.
Upstream didn’t want to review or include them in tree if there were no users and also wouldn’t have accepted me putting my CUDA changes in the PR introducing those hooks so i gave up 🤷♂️ | 17:01:58 |
| connor (he/him) | The CUDA compat path and driver link lib path should appear before any stub directory entries in the runpath… but we also shouldn’t have any stub directory entries in the runpath | 17:03:24 |
| connor (he/him) | The other approach I took was to register a function, which registers all the other functions present in the setup hook, to be executed during prePhases.
Since that’s an insane thing to do, it ensures that our hooks are the last in the order of every phase we’re adding them to.
https://github.com/ConnorBaker/cuda-packages/blob/8a317116a07717b13e0608f47b78bd6d75f8bb99/pkgs/development/cuda-modules/packages/nvccHook/nvccHook.bash#L24
| 17:10:55 |
| connor (he/him) | Part of the reason I did that was because the order Nixpkgs loads dependencies is wrong (there’s an issue from Ericson from like 8y ago)
As a result, even though my setup hooks depend on the array utility functions, they’re not available when my setup hooks is sourced — only after the phases have begun | 17:13:35 |
