| 24 Nov 2025 |
 yorik.sar | Sure, I could join. | 17:00:21 |
 SomeoneSerge (back on matrix) | connor (burnt/out) (UTC-8) Gaétan Lepage doing same time as last week right? | 17:08:59 |
| SomeoneSerge (back on matrix) | * connor (burnt/out) (UTC-8) Gaétan Lepage doing same time as last week right? (sent an update, Connor yours seems to auto-reject) | 17:09:47 |
 connor (burnt/out) (UTC-8) | Huh, I don’t even see anything in my junk mail :/ | 17:31:54 |
| connor (burnt/out) (UTC-8) | Two thoughts:
- Helion backend for einops
- Optuna integration with Helion to allow for persistent studies (e.g., if the seed is fixed and the number of generations is increased, the optimization should resume from where it stopped rather than start an entirely new optimization)
| 17:34:43 |
| connor (burnt/out) (UTC-8) | Serge I’m working on a matrix of behavior of propagating and consuming setup hooks through packages | 17:36:11 |
| SomeoneSerge (back on matrix) | I think the key bit is that buildInputs already are (0, 0) away from the current derivation (because we're building for (0, 1)) | 17:59:02 |
| SomeoneSerge (back on matrix) | But for the PR we should just drop the offset-checking logic from the hook imo | 18:00:16 |
| SomeoneSerge (back on matrix) | * But for the PR we should just drop the offset-checking logic from the hook imo, and avoid propagating extra patchelf | 18:00:45 |
| connor (burnt/out) (UTC-8) | Created this table by building everything in the scope I added and assembling the information from the logs (no build succeeds since I don't produce an out output): https://github.com/ConnorBaker/nix-propagation-behavior/blob/1afbd58f2af1468d4564722b7180cc4d89967ef3/README.md | 19:28:30 |
| connor (burnt/out) (UTC-8) | Using the table, our "users" are packages consuming the stub outputs, which we would expect to happen from nativeBuildInputs or buildInputs (or both).
If used in nativeBuildInputs, we need to install the hook in propagated-host-host-deps (for (-1, -1)) or propagated-build-inputs (for (-1, 0)).
If used in buildInputs, we need to install the hook in propagated-build-build-deps (for (-1, -1)) or propagated-native-build-inputs (for (-1, 0)).
Since we're modifying binaries in-place through patchelf, the (-1, -1) offsets make more sense to me, but I can install to propagated-build-inputs and propagated-native-build-inputs instead for the (-1, 0) offsets. | 19:28:35 |
| connor (burnt/out) (UTC-8) | I think the "wonky" behavior I had seen previously (essentially the hook wasn't sourced when put in propagated-build-inputs or propagated-native-build-inputs) can be explained by the fact I was installing the hook during postInstall. Since the dependency files are replaced in fixupPhase, those entries were clobbered.
Now that I'm installing them in postFixup, it's not an issue. | 19:30:10 |
| connor (burnt/out) (UTC-8) | https://github.com/NixOS/nixpkgs/pull/459416 has been updated, please consider merging it | 21:35:10 |
| SomeoneSerge (back on matrix) | Nice, thanks! I had one like this somewhere too, clearly haven't looked at it in too long
But yeah must be (( "$relHostOffset" <= "$relTargetOffset" )) || continue
| 22:32:54 |
| 25 Nov 2025 |
| connor (burnt/out) (UTC-8) | https://github.com/NixOS/nixpkgs/pull/464779 | 02:37:33 |
| connor (burnt/out) (UTC-8) | hope to have tensorrt version bump up tomorrow... but that also involves bumping onnx-tensorrt, tensorrt-oss, and a few other things simultaneously :F | 04:45:45 |
| connor (burnt/out) (UTC-8) | while building every combination of TensorRT 10.x and CUDA release on x86_64-linux to verify I could remove cudnn and nvrtc as buildInputs I prodded Claude to make the optuna+helion thing I mentioned earlier: https://github.com/ConnorBaker/helion/blob/claude/enhance-helion-autotuner-01V2JcV61tYZPi7y6cuZpE54/helion/autotuner/optuna_search.py
It doesn't work with the version of Helion we have in Nixpkgs because of some API changes to parallel_benchmarking but it seems like what I wanted. Haven't had the chance to mess with it though :l | 04:50:08 |
| connor (burnt/out) (UTC-8) | https://github.com/NixOS/nixpkgs/pull/464947 | 15:01:41 |
| connor (burnt/out) (UTC-8) | https://github.com/NixOS/nixpkgs/pull/464957 | 15:52:14 |
| connor (burnt/out) (UTC-8) | https://github.com/NixOS/nixpkgs/pull/465047 | 20:23:29 |
 Gaétan Lepage | https://github.com/NixOS/nixpkgs/pull/450587 | 20:24:55 |
| 26 Nov 2025 |
| connor (burnt/out) (UTC-8) | SomeoneSerge (back on matrix) I updated https://github.com/NixOS/nixpkgs/pull/459416 please merge | 02:57:01 |
| connor (burnt/out) (UTC-8) | 🤦♂️ | 15:37:52 |
| connor (burnt/out) (UTC-8) | https://github.com/onnx/onnx-tensorrt/pull/1043 moved from the pycuda to cuda python package and documented it nowhere I've been able to find | 15:38:34 |
| connor (burnt/out) (UTC-8) | Kevin Mittman (UTC-7): the entry for --verbose seems to be for --no-dry-run, which doesn't have an entry: https://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/#verbose-v | 18:12:52 |
| connor (burnt/out) (UTC-8) | well, pycuda's tests exposed a few issues with how NVCC works
notably that it dies if GCC isn't on the path -- it doesn't even use the path modifications from nvcc.profile | 18:38:12 |
| connor (burnt/out) (UTC-8) | so this is now a larger PR https://github.com/NixOS/nixpkgs/pull/465047 | 18:38:20 |
| connor (burnt/out) (UTC-8) | I'm going to split out the NVCC changes to a different PR and make pycuda a stacked PR | 18:51:51 |
| connor (burnt/out) (UTC-8) | :F | 18:51:53 |
| connor (burnt/out) (UTC-8) | https://github.com/NixOS/nixpkgs/pull/465329 | 18:57:02 |
