🔥 PyTorch 2.10.0 (+ triton 3.6.0) 🔥

Changelog: https://github.com/pytorch/pytorch/releases/tag/v2.10.0
PR: https://github.com/NixOS/nixpkgs/pull/484881
PR tracker: https://nixpkgs-tracker.ocfox.me/?pr=484881

This one took quite a while to bring to nixpkgs, and I'm glad to have finally gotten it merged!

Some basic testing has been done (basic builds, cudaSupport builds, some gpuChecks).
However, exhaustively building and testing all the downstream dependencies isn't feasible (at least without more time and hardware).
-> Please, don't hesitate to report any breakage in this channel, and feel free to ping me as well.

Thanks a lot to everyone who helped, and more generally to everyone else for your patience.

Related to PyTorch but not to 2.10, it looks like my CI cannot build it since yesterday morning (it was still python3.13-torch-2.9.1) 🤔
Torch 2.10.0 does not solve that issue.

I build with:

      run: "cd nixpkgs-master\n  nix-build -I nixpkgs=. --arg config '{ allowUnfree\
        \ = true; cudaSupport = true; openclSupport = true; rocmSupport = false;}'\
        \ --option allow-import-from-derivation false --max-jobs 1 -A python313Packages.torch"

Here is the last part of my build log (after 3h12 of build).
https://pad.lassul.us/s/xWhnOmbeo2#

Any idea what could be the cause?

Related to PyTorch but not to 2.10, it looks like my CI cannot build it since yesterday morning (it was still python3.13-torch-2.9.1) 🤔
Torch 2.10.0 does not solve that issue.

I build with:

      run: "cd nixpkgs-master\n  nix-build -I nixpkgs=. --arg config '{ allowUnfree\
        \ = true; cudaSupport = true; openclSupport = true; rocmSupport = false;}'\
        \ --option allow-import-from-derivation false --max-jobs 1 -A python313Packages.torch"

Here is the last log (after 3h12 of build).
https://bpa.st/5CQA

Any idea what could be the cause?

Hi! Thanks :)
I rebased the PR yesterday: https://github.com/NixOS/nixpkgs/pull/490175

I couldn't do much as its dependendency xgrammar is broken. Once I'll properly bump the {c,q}utlass et al. overrides, then sure, help with testing will be welcome :)

RE: vllm update to 0.16.0

Although vllm 0.16.0 was initially depending on torch==2.10.0, this changed.
Indeed, upstream force-pushed the tag to revert this change and go back to requiring torch==2.9.1.

I learnt about this by chance, discussing this wis a coworker.

https://github.com/nixos-cuda/infra/commit/eeb5eb95d8eb0abba5fa50d14500c0d20a2e5d12

🥲