| 1 Nov 2025 |
 hexa | shouldn't that be propagatedNativeBuildInputs? | 02:10:31 |
 connor (he/him) | For backendStdenv.cc in cuda_nvcc.nix or for setupCudaHook in buildRedist/default.nix? | 02:11:19 |
| hexa | cuda_nvcc.nix | 02:12:14 |
 SomeoneSerge (back on matrix) | Here's the old logic: https://github.com/NixOS/nixpkgs/blob/99fe7bea35ed3ca8de66188d2337cb3e7c6e83e7/pkgs/development/cuda-modules/_cuda/fixups/cuda_nvcc.nix#L50 | 02:12:30 |
| hexa | it's too late here, going to sleep | 02:13:19 |
| connor (he/him) | backendStdenv.cc needs to be in propagatedBuildInputs; makeSetupHook does the same thing | 02:13:21 |
| SomeoneSerge (back on matrix) | We do reference the host compiler from nvcc and from the hook, but we do not propagate | 02:13:46 |
| SomeoneSerge (back on matrix) | Tbh the hook referencing the compiler looks sus too, but my memory's all gone | 02:16:08 |
| SomeoneSerge (back on matrix) | Ah! I see now why the hook was "ok": it would only propagate stuff if cudaPropagateToOutput is explicitly set by the downstream derivation | 02:21:42 |
| SomeoneSerge (back on matrix) | Ok but the onnxruntime part I didn't look at/don't know don't understand | 02:23:57 |
| SomeoneSerge (back on matrix) | My best guess is it's because an extraneous propagatedBuildInputs = [ setupCudaHook ] also slipped in into buildRedist.nix, but tbh I'm still struggling to explain how this leads to the extra reference | 02:51:56 |
| SomeoneSerge (back on matrix) | Another person on gh saying they found a legit nvcc reference in libonnxruntime_providers_cuda, so I think this can't be explained by hooks alone | 02:56:09 |
| connor (he/him) | Is it possible onnxruntime does JIT compilation | 02:57:40 |
| SomeoneSerge (back on matrix) | Yeah that what I'm wondering | 03:10:52 |
| SomeoneSerge (back on matrix) | Nope | 03:12:13 |
| SomeoneSerge (back on matrix) | More mysteries! https://github.com/NixOS/nixpkgs/pull/457424#issuecomment-3475537174 | 03:15:11 |
| SomeoneSerge (back on matrix) | connor (burnt/out) (UTC-7): when you started propagating crt/host_config.h from cudart, did you also drop the manual nvcc from any leaf packages? Figure you might be quicker to answer than me reading or running tests | 03:58:51 |
| connor (he/him) | Manual NVCC? | 04:00:59 |
| SomeoneSerge (back on matrix) | Like we had buildInputs = [ nvcc ] in a bunch of places because of the host_config.h dependency | 04:03:40 |
| SomeoneSerge (back on matrix) | Which will not have been needed if cudart propagates it | 04:03:57 |
| connor (he/him) | I don’t remember honestly
I think I also had it propagate because headers need to be in buildInputs to be discovered | 04:05:08 |
| connor (he/him) | I just looked through the commits for the CUDA 13 PR; I didn't see removal of nvcc from buildInputs anywhere | 04:13:19 |
| SomeoneSerge (back on matrix) | Aight this may well be my favourite Halloween story so far, and still open-ended: https://github.com/NixOS/nixpkgs/pull/457424#issuecomment-3475742510 | 04:17:14 |
| SomeoneSerge (back on matrix) | I can't explain this | 04:19:45 |
| SomeoneSerge (back on matrix) | No way | 04:19:48 |
| connor (he/him) | can you imagine if it's non-deterministic | 04:24:09 |
| connor (he/him) | or only occurs if there are multiple capabilities | 04:24:16 |
| connor (he/him) | I mean I was able to reproduce with a single capability | 04:24:54 |
| SomeoneSerge (back on matrix) | Yeah. If I were to try to explain it, I'd start checking if they e.g. try to copy CUDAToolkit_INCLUDE_DIRS or any other FindCUDAToolkit's targets' properties into their own targets | 04:31:09 |
| connor (he/him) | For what it's worth, I looked through the tree and here are the packages I found where cuda_nvcc ends up in buildInputs in some way. This list doesn't include other way it could end up in the closure -- string interpolation, environment variables, cmakeFlags, etc.
- ./pkgs/by-name/ba/basalt-monado/package.nix
- ./pkgs/by-name/dl/dlib/package.nix
- ./pkgs/by-name/fr/frei0r/package.nix
- ./pkgs/by-name/gp/gpu-burn/package.nix
- ./pkgs/by-name/ka/katago/package.nix
- ./pkgs/by-name/ko/koboldcpp/package.nix
- ./pkgs/by-name/ma/mathematica/generic.nix
- ./pkgs/by-name/mf/mfaktc/package.nix
- ./pkgs/by-name/mo/monado/package.nix
- ./pkgs/by-name/xm/xmrig-cuda/package.nix
- ./pkgs/by-name/xm/xmrig-cuda-mo/package.nix
- ./pkgs/development/cuda-modules/packages/cuda_cudart.nix
- ./pkgs/development/cuda-modules/packages/nccl.nix
- ./pkgs/development/libraries/ffmpeg/generic.nix
- ./pkgs/development/python-modules/causal-conv1d/default.nix
- ./pkgs/development/python-modules/cupy/default.nix
- ./pkgs/development/python-modules/jaxlib/default.nix
- ./pkgs/development/python-modules/lightgbm/default.nix
- ./pkgs/development/python-modules/mamba-ssm/default.nix
- ./pkgs/development/python-modules/tensorflow/default.nix
- ./pkgs/development/python-modules/torch/source/default.nix
- ./pkgs/development/python-modules/warp-lang/default.nix
| 04:32:50 |
