Help-Wanted Priority List

hmaarrfk commented 1 week ago

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
https://github.com/conda-forge/pytorch-cpu-feedstock/issues/274

Maintainability

Consider Upstream's split build: https://github.com/conda-forge/pytorch-cpu-feedstock/issues/272
Unvendoring: https://github.com/conda-forge/pytorch-cpu-feedstock/issues/108
cuDSS: https://github.com/conda-forge/pytorch-cpu-feedstock/issues/269

User support

Feature addition

Triton support: https://github.com/conda-forge/pytorch-cpu-feedstock/issues/166
MKLDNN + CUDA: https://github.com/conda-forge/pytorch-cpu-feedstock/issues/233
Other blas + CUDA: https://github.com/conda-forge/pytorch-cpu-feedstock/issues/177
GDS: https://github.com/conda-forge/pytorch-cpu-feedstock/issues/257
https://github.com/conda-forge/pytorch-cpu-feedstock/issues/269
Optimized builds CUDA 12.2 or 12.4 https://github.com/conda-forge/pytorch-cpu-feedstock/issues/270
Identifying the broken builds OSX arm + Pytorch https://github.com/conda-forge/pytorch-cpu-feedstock/issues/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 -- https://github.com/conda-forge/pytorch-cpu-feedstock/issues/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 https://github.com/conda-forge/pytorch-cpu-feedstock/pull/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