Help-Wanted Priority List

[hmaarrfk]

By no means a strict order but here is my opinion on where we can use the most help. Jump in where you can.

While feature additions are helpful, we prioritize maintainability and compatibility with the rest of the conda-forge ecosystem.

Please try to discuss implementation details in the dedicated issue or PR.

Unix Stream

Maintenance

• "RC Builds": current version is 2.5.0rcX

Maintainability

• Consider Upstream's split build: Use upstream's new split build #272
• Unvendoring: Splitting this package in managable chunks #108
• cuDSS: cuDSS support in 2.5.0(?) #269

User support

• Enable linux ppc64le #255
• Make Magma optional for cuda builds? #275

Feature addition

• Triton support: building with triton support?  #166
• MKLDNN + CUDA: Build with mkldnn support alongside cuda #233
• Other blas + CUDA: Enable other BLAS implementations with CUDA #177
• GDS: (Future release) building with GDS #257
• cuDSS support in 2.5.0(?) #269
• Optimized builds CUDA 12.2 or 12.4 Building with CUDA 12.2 #270
• Identifying the broken builds OSX arm + Pytorch Segfault when importing sklearn as well, osx-arm64 #260

Windows Stream

Mostly this needs a dedicated champion for the Windows cause. The large builds make this incompatible with the conda-forge philosophy of building in the open using automated bots.

• A powerful windows CI for >6 hr builds -- Large Windows Runners #268
  • May be substituted with a benevolent contributor.
  • Contributor must be responsive and adhere to CFEP03
• Windows CPU + system ABSL
  • A proof of concept Windows build was demonstrated in 2024 in feat: yet another attempt to add windows builds #231 We just prefer to use conda-forge ABSL instead of rebuilding it for each feedstock.
  • If a maintainer requests CPU builds for themselves, and is unable to host on their own channel, then this might be acceptable as getting merged.
• Windows GPU Builds
  • Challenging to fit within 6 hrs.
  • Some cuda architectures may be dropped

Completed

• Pytorch 2.4.1 - Pytorch 2.4.1, python 3.13, cudnn9 #261
• CUDNN 9 - Pytorch 2.4.1, python 3.13, cudnn9 #261
• ABSEIL 20240722 migration Rebuild for libabseil 20240722, libgrp 1.65 & libprotobuf 5.27.5 #265
• Missing runtime dependency on setuptools #274

[atalman]

Hi @hmaarrfk what is the point about Maintenance: "RC Builds": current version is 2.5.0rcX ?

[hmaarrfk]

Hi @hmaarrfk what is the point about Maintenance: "RC Builds": current version is 2.5.0rcX ?

Generally about staying ahead of any compilation issues and being able to report small patches to help cross compilation and whatnot.

[carterbox]

I've opened a PR for #274; it's #276

[hmaarrfk]

@jeongseok-meta @jslee02 if you are interested in going through this list. The separation to use upstream's libtorch's mechanism specifyign the flags:

#   BUILD_LIBTORCH_WHL
#      Builds libtorch.so and its dependencies as a wheel
#
#   BUILD_PYTHON_ONLY
#      Builds pytorch as a wheel using libtorch.so from a seperate wheel

https://github.com/pytorch/pytorch/blob/d84a34441068abc46fe944dea39d4cc8026b824b/setup.py#L213
as needed.

CFEP03 got us VERY far, but it really requires a dedicated champion (which I know is hard to guarantee as priorities change), but if you have access to access to machines, then you can help!