libtorch-bin: 2.0.0 -> 2.2.2

[pinpox]

Description of changes

Also fixed the prefetch script, the url seems to have slightly changed Fixes #299311

Things done

• Built on platform(s)
  • x86_64-linux
  • aarch64-linux
  • x86_64-darwin
  • aarch64-darwin
• For non-Linux: Is sandboxing enabled in nix.conf? (See Nix manual)
  • sandbox = relaxed
  • sandbox = true
• Tested, as applicable:
  • NixOS test(s) (look inside nixos/tests)
  • and/or package tests
  • or, for functions and "core" functionality, tests in lib/tests or pkgs/test
  • made sure NixOS tests are linked to the relevant packages
• Tested compilation of all packages that depend on this change using nix-shell -p nixpkgs-review --run "nixpkgs-review rev HEAD". Note: all changes have to be committed, also see nixpkgs-review usage
• Tested basic functionality of all binary files (usually in ./result/bin/)
• 24.05 Release Notes (or backporting 23.05 and 23.11 Release notes)
  • (Package updates) Added a release notes entry if the change is major or breaking
  • (Module updates) Added a release notes entry if the change is significant
  • (Module addition) Added a release notes entry if adding a new NixOS module
• Fits CONTRIBUTING.md.

---

Add a 👍 reaction to pull requests you find important.

[msanft]

Result of nixpkgs-review pr 302427 run on x86_64-linux 1

1 package failed to build:

• ocamlPackages.torch

3 packages built:

• in-formant
• libtorch-bin
• libtorch-bin.dev

[pinpox]

1 package failed to build:

* ocamlPackages.torch

Hm, not sure wahat to do about the ocaml package. @bcdarwin could you take a look?

[junjihashimoto]

@pinpox Thank you for updating it.
Could you change cu118 to cu121? Because pytorch's default(pip) is cu121.

[pinpox]

Could you change cu118 to cu121? Because pytorch's default(pip) is cu121.

done.

[bcdarwin]

1 package failed to build:

* ocamlPackages.torch

Hm, not sure wahat to do about the ocaml package. @bcdarwin could you take a look?

Ocaml-torch hasn't been updated in a year or so, so not surprising it would eventually break. Happy for you to mark broken for now.

[npatsakula]

Hello @pinpox!

I've tested new version with hash from this PR on x86_64 (linux + CPU), so maybe we can mark checkbox as completed:
#302184

Also I need to mention that 2.2.2 has support of aarch64 Darwin and it would be great to add that to this PR.

@junjihashimoto, it seems that I need to close my PR as a duplicate.

[junjihashimoto]

@npatsakula
Thank you for pointing it out.
It is preferable that the new version also uses cu121 not cu118.

[pinpox]

aarch64 added (thanks @npatsakula !). Let me know if something else is missing.

[junjihashimoto]

LGTM!

[msanft]

Result of nixpkgs-review pr 302427 run on x86_64-linux 1

3 packages built:

• in-formant
• libtorch-bin
• libtorch-bin.dev

[pbsds]

Please squash the commits, and could you look into the aarch64-darwin issue?

[pinpox]

Please squash the commits, and could you look into the aarch64-darwin issue?

fixed and squashed

[junjihashimoto]

Thx!

[pbsds]

aarch64-darwin is still failing, did you forget to push the change?

[npatsakula]

aarch64-darwin is still failing, did you forget to push the change?

Hello @pbsds! My bad, looks like I need your help:

1. I don't understand this error message and I am unable to associate it to particular package test.
2. I can't reproduce this test on my aarch64-darwin machine, It compiles without errors (I'm using this package from flake).

[pbsds]

I'm not familiar with nor have access to darwin, but as I understand the darwin build sandbox is lacking. Perhaps it passes on your machine since the dyld check is able to match with a dylib on your system that the ofborg builders don't have installed?

Comparing the fixupPhase output in the full logs: https://logs.ofborg.org/?attempt_id=e65c3bce-5f6e-450f-9be7-f390bfbb6790&key=nixos%2Fnixpkgs.302427

x86_64-darwin:

/nix/store/h7rmvcdf3vlizf0nzjadxy18vh60ngin-libtorch-2.2.2/lib/libtorch_python.dylib: fixing dylib
/nix/store/h7rmvcdf3vlizf0nzjadxy18vh60ngin-libtorch-2.2.2/lib/libbackend_with_compiler.dylib: fixing dylib
/nix/store/h7rmvcdf3vlizf0nzjadxy18vh60ngin-libtorch-2.2.2/lib/libtorch.dylib: fixing dylib
/nix/store/h7rmvcdf3vlizf0nzjadxy18vh60ngin-libtorch-2.2.2/lib/libtorch_global_deps.dylib: fixing dylib
/nix/store/h7rmvcdf3vlizf0nzjadxy18vh60ngin-libtorch-2.2.2/lib/libiomp5.dylib: fixing dylib
/nix/store/h7rmvcdf3vlizf0nzjadxy18vh60ngin-libtorch-2.2.2/lib/libtorch_cpu.dylib: fixing dylib
/nix/store/h7rmvcdf3vlizf0nzjadxy18vh60ngin-libtorch-2.2.2/lib/libjitbackend_test.dylib: fixing dylib
/nix/store/h7rmvcdf3vlizf0nzjadxy18vh60ngin-libtorch-2.2.2/lib/libc10.dylib: fixing dylib
/nix/store/h7rmvcdf3vlizf0nzjadxy18vh60ngin-libtorch-2.2.2/lib/libtorchbind_test.dylib: fixing dylib
/nix/store/h7rmvcdf3vlizf0nzjadxy18vh60ngin-libtorch-2.2.2/lib/libshm.dylib: fixing dylib

aarch64-darwin:

/nix/store/2cfbhxddvg39fc6ldqk0kk7qs7vgrigz-libtorch-2.2.2/lib/libtorch_python.dylib: fixing dylib
/nix/store/2cfbhxddvg39fc6ldqk0kk7qs7vgrigz-libtorch-2.2.2/lib/libtorch.dylib: fixing dylib
/nix/store/2cfbhxddvg39fc6ldqk0kk7qs7vgrigz-libtorch-2.2.2/lib/libtorch_global_deps.dylib: fixing dylib
/nix/store/2cfbhxddvg39fc6ldqk0kk7qs7vgrigz-libtorch-2.2.2/lib/libtorch_cpu.dylib: fixing dylib
/nix/store/2cfbhxddvg39fc6ldqk0kk7qs7vgrigz-libtorch-2.2.2/lib/libc10.dylib: fixing dylib
/nix/store/2cfbhxddvg39fc6ldqk0kk7qs7vgrigz-libtorch-2.2.2/lib/libshm.dylib: fixing dylib

Some seem to be missing

[junjihashimoto]

nixpkgs/pkgs/development/libraries/science/math/libtorch/bin.nix

Line 24 in dd25cfd

libcxx-for-libtorch = if stdenv.hostPlatform.system == "x86_64-darwin" then libcxx else stdenv.cc.cc.lib;

Could you update the code to use stdenv.isDarwin?

[junjihashimoto]

Thx!

[pbsds]

builder failure... I'd love to verify the darwin fix

@ofborg build libtorch-bin

[junjihashimoto]

@pinpox @pbsds

Probably platforms should be updated like this PR.

    platforms = with platforms; [
      "aarch64-darwin"
      "x86_64-darwin" "x86_64-linux"
    ];

https://github.com/pinpox/nixpkgs/pull/2/files#diff-3a9f32ef2e590562647c6c78cd8af994c4ef85795f4ef1744e5d43069bdc3155R102-R105

[Aleksanaa]

Yes, keep in mind Unix is not Linux + Darwin and all is not x86_64 + aarch64

[pbsds]

It worked before the push, and now it failed again.
The failure seems to be a lack of libomp.dylib, which i assume is builder dependent.
Pytorch darwin assumes libomp (pytorch/pytorch#20030) which should be provided by llvmPackages.openmp. If the binaries are patched with this we should get a more hermetic runtime closure

[wegank]

Superseded by #308328.