FYI: The PR updating jax to 0.9.2 and the rest of the ecosystem with it (22 commits 🫠) has finally been merged! (after 2 months of efforts)

Please let me know if you notice any regression.

Changelogs:

• https://github.com/jax-ml/jax/releases/tag/jax-v0.9.0
• https://github.com/jax-ml/jax/releases/tag/jax-v0.9.1
• https://github.com/jax-ml/jax/releases/tag/jax-v0.9.2

https://github.com/NixOS/nixpkgs/pull/482118

Looking for reviewers on:

• https://github.com/NixOS/nixpkgs/pull/503855
• https://github.com/NixOS/nixpkgs/pull/502810

🔥 PyTorch 2.11.0 (+ torchcodec 0.11.0) 🔥

Changelog: https://github.com/pytorch/pytorch/releases/tag/v2.11.0
PR: https://github.com/NixOS/nixpkgs/pull/502810
PR tracker: https://nixpkgs-tracker.ocfox.me/?pr=502810

Notable changes on the nix packaging side:

• rocmSupport for torchcodec
• rocmSupport for torchaudio
• revamp of the torchaudio derivation, addition of testing
• Addition of torchaudio.gpuCheck

This one only took me 4 days 🚀 I hope it doesn't break too many things 🥸

In reply to @ccicnce113424:matrix.org
https://github.com/NixOS/nixpkgs/pull/498612

https://github.com/NixOS/nixpkgs/pull/505958

• I need to support onnx-tensorrt
• It's an absolute pain in the ass to package and make work over all the CUDA versions I have to support (cough cuda-legacy cough)
• It's not packaged in-tree (I packaged a (then new!) copy here https://github.com/ConnorBaker/cuda-packages/blob/7604ebdb8e9484c633710b408e50816e95ebfac9/pkgs/by-name/on/onnx-tensorrt/package.nix)
• I don't understand why ONNX_ML being set to 1 broke onnx-tensorrt or even know if it's still a problem

In reply to @ccicnce113424:matrix.org
ping

Unrelated, but I'm getting errors with the nix-required-mounts hook, maybe I'm just on a bad commit:

$ /nix/store/kgcbq3ablba98myqn8j4sq7yla6nzs3m-nix-required-mounts-0.0.1/bin/nix-required-mounts /nix/store/5i5w8byychlxjbrjnvfl4rwbi9wqr66d-nix-shell-env.drv
Traceback (most recent call last):
  File "/nix/store/kgcbq3ablba98myqn8j4sq7yla6nzs3m-nix-required-mounts-0.0.1/bin/.nix-required-mounts-wrapped", line 9, in <module>
    sys.exit(entrypoint())
             ~~~~~~~~~~^^
  File "/nix/store/kgcbq3ablba98myqn8j4sq7yla6nzs3m-nix-required-mounts-0.0.1/lib/python3.13/site-packages/nix_required_mounts.py", line 142, in entrypoint
    [canon_drv_path] = parsed_drv.keys()
    ^^^^^^^^^^^^^^^^
ValueError: too many values to unpack (expected 1)

Using commit https://github.com/NixOS/nixpkgs/commits/8110df5ad7abf5d4c0f6fb0f8f978390e77f9685 of Nixpkgs. I remember something about JSON derivation format change but I can't remember for the life of me if we already fixed that. Previous commit was https://github.com/NixOS/nixpkgs/commits/a6531044f6d0bef691ea18d4d4ce44d0daa6e816.

Interesting, I'm facing another error:

error:
       … while setting up the build environment

       error: getting attributes of path '/nix/store/02wkfkv635277dyq176lw5dcqpxlpsl0-kmod-31/sbin/bin': No such file or directory

In reply to @connorbaker:matrix.org
Also, can someone explain https://developer.download.nvidia.com/compute/cuda/redist/cuda_compat_orin/ to me? If I had to hazard a guess, it would be that it would allow using CUDA 13.1/13.2 on a Jetson based on JetPack 7 (since support for Orin is added in JetPack 7.2, but that hasn't been released yet as far as I can tell, though it is talked about in the CUDA 13.2 blog post: https://developer.nvidia.com/blog/cuda-13-2-introduces-enhanced-cuda-tile-support-and-new-python-features/#embedded_devices). Is that correct? Or does it allow forward compat across major versions and can be used with JetPack 6, similar to how cuda_compat for Orins on JetPack 5 allows support for up to CUDA 12.2 despite shipping with CUDA 11.4.

In reply to @connorbaker:matrix.org
Also, can someone explain https://developer.download.nvidia.com/compute/cuda/redist/cuda_compat_orin/ to me? If I had to hazard a guess, it would be that it would allow using CUDA 13.1/13.2 on a Jetson based on JetPack 7 (since support for Orin is added in JetPack 7.2, but that hasn't been released yet as far as I can tell, though it is talked about in the CUDA 13.2 blog post: https://developer.nvidia.com/blog/cuda-13-2-introduces-enhanced-cuda-tile-support-and-new-python-features/#embedded_devices). Is that correct? Or does it allow forward compat across major versions and can be used with JetPack 6, similar to how cuda_compat for Orins on JetPack 5 allows support for up to CUDA 12.2 despite shipping with CUDA 11.4.

Neat project!
https://www.youtube.com/watch?v=AvK_gi_snJE

Some quotes from the conclusions:

"Nixpkgs CUDA aarch64 ecosystem is quite healthy"
Thanks to nix-community's 80 core ampere system

"Flakes make a good user interface"
This one is for you SomeoneSerge (matrix works sometimes)

"Lack of cached aarch64 CUDA builds"
👀💸

"Uneven freshness of packages"
Until we unlock core and maintainer cloning, that's going to be tough to change (especially reagrding how fast the whole space is moving).

"Community lacks GPU hardware"
Yes. And CPU too 😅