!eWOErHSaiddIbsUNsJ:nixos.org

NixOS CUDA

282 Members
CUDA packages maintenance and support in nixpkgs | https://github.com/orgs/NixOS/projects/27/ | https://nixos.org/manual/nixpkgs/unstable/#cuda58 Servers

Load older messages


SenderMessageTime
1 Dec 2025
@connorbaker:matrix.orgconnor (burnt/out) (UTC-8)Make sure you’re setting config.cudaCapabilities explicitly to speed up compilation12:39:34
@connorbaker:matrix.orgconnor (burnt/out) (UTC-8)

Packaging onnx/onnx-tensorrt/onnxruntime and getting them working for a range of cuda/cudnn/tensorrt releases is maddening oh my god:

  • Fix for GCC 11/NVCC's template processing https://github.com/nixos-cuda/cuda-legacy/commit/79c3026f1a761125b6faa609c5f9e56e723a4568
  • https://github.com/nixos-cuda/cuda-legacy/issues/6 since ONNX Runtime needs GCC > 11.1 and I need it to build with 11.4
  • I'm fairly certain I managed to get a build failure from ONNX Runtime with CUDA 11.8 because a reference to cuda_nvcc was held onto, so ONNX Runtime failed the build
21:32:54
2 Dec 2025
@connorbaker:matrix.orgconnor (burnt/out) (UTC-8)Okay NVCC from CUDA 11.4 still dies horrifically when using newer versions of glibc with GCC 11 so that’ll stay on GCC 10. Also it seems onnxruntime 1.22.2 needs CCCL newer than what 11.4 provides, so supporting onnxruntime on 11.4 means finding an older version which works :l01:21:58
@glepage:matrix.orgGaétan Lepage connor (burnt/out) (UTC-8) I approved your onnx PR 08:14:41
@glepage:matrix.orgGaétan LepageI can tackle the bump to 1.20.0 after it's merged08:14:56
@keiichi:matrix.orgtetois caching disabled for unstable ? I have to build onnxruntime and then I read this https://discourse.nixos.org/t/cuda-cache-for-nix-community/56038/3 11:45:21
@hexa:lossy.networkhexa (UTC+1)building cuda packages on nix-community was discontinued12:07:37
@hexa:lossy.networkhexa (UTC+1)there is now the flox cache, which builds for their versions of nixpkgs12:08:00
@hexa:lossy.networkhexa (UTC+1)* there is now the flox cache, which builds for their versions of nixpkgs, which don't really align with our channels12:08:09
@hexa:lossy.networkhexa (UTC+1)and possibly the build results from nixos-cuda.org12:08:32
@keiichi:matrix.orgtetoI had totally missed that. The discourse comment made it sound like it is retaliation against flox actions ?12:11:30
@hexa:lossy.networkhexa (UTC+1)Redacted or Malformed Event12:45:08
@connorbaker:matrix.orgconnor (burnt/out) (UTC-8)I am 🔪23:20:33
@connorbaker:matrix.orgconnor (burnt/out) (UTC-8)

okay how the hell did you all manage to debug this on master:

error: output '/nix/store/pb5lmj07b9vs4awlxrga9s6wkfzzrk1i-onnxruntime-1.22.2-dev' is not allowed to refer to the following paths:
         /nix/store/rf7a79c1a73ny3j6jw77nza4wd5z40i1-cuda11.4-cuda_nvcc-11.4.152
23:20:49
@connorbaker:matrix.orgconnor (burnt/out) (UTC-8)(I am building ONNX Runtime for CUDA 11.4 for a single architecture)23:21:13
@connorbaker:matrix.orgconnor (burnt/out) (UTC-8) CC Gaétan Lepage SomeoneSerge (back on matrix) 23:21:23
3 Dec 2025
@connorbaker:matrix.orgconnor (burnt/out) (UTC-8) ugh
it's through cudnn-frontend
because https://github.com/NVIDIA/cudnn-frontend/blob/0258951d4d512f4714eb1574496f4d57669b1b93/CMakeLists.txt#L43 means the generate cmake target has the include directory of NVCC
and since I made NVCC a single-output derivation, the output containing include and bin are one and the same
so ONNX Runtime's disallowedRequisites = lib.optionals cudaSupport [ (lib.getBin cuda_nvcc) ]; is tripped because of the cmake config from cudnn-frontend
so I guess I'm trying to split up NVCC again
fml
00:09:44
@glepage:matrix.orgGaétan Lepage FYI, https://hydra.nixos-cuda.org is now building nixos-25.11-small too :) 09:50:51
@glepage:matrix.orgGaétan Lepage * FYI (for stable channel users), https://hydra.nixos-cuda.org is now building nixos-25.11-small too :) 09:51:06
@hexa:lossy.networkhexa (UTC+1)after renaming you need to toggle the jobsets for them to work again09:52:00
@hexa:lossy.networkhexa (UTC+1)otherwise they just09:52:10
@hexa:lossy.networkhexa (UTC+1)

evaluation failed with exit code 255

09:52:15
@hexa:lossy.networkhexa (UTC+1)as can be seen here https://hydra.nixos-cuda.org/jobset/nixos-cuda/cuda-packages-unstable#tabs-errors09:52:24
@hexa:lossy.networkhexa (UTC+1)and here https://hydra.nixos-cuda.org/jobset/nixos-cuda/cuda-packages-legacy#tabs-errors09:52:38
@hexa:lossy.networkhexa (UTC+1)https://github.com/NixOS/hydra/issues/128809:52:56
@glepage:matrix.orgGaétan Lepage Thanks for the tip hexa (UTC+1)! 09:54:14
@hexa:lossy.networkhexa (UTC+1) Gaétan Lepage: you probably also want a lower priority (higher value) than cache.nixos.org 15:58:01
@hexa:lossy.networkhexa (UTC+1) * Gaétan Lepage: you probably also want a lower priority (higher value) than cache.nixos.org, which has prio 30 15:58:10
@glepage:matrix.orgGaétan LepageWhat are you referring to? Something in the nix-cuda infra?17:36:09
@hexa:lossy.networkhexa (UTC+1)https://cache.nixos-cuda.org/nix-cache-info18:22:39

Show newer messages


Back to Room ListRoom Version: 9