add deepspeed

[hadim]

Checklist

• Title of this PR is meaningful: e.g. "Adding my_nifty_package", not "updated meta.yaml".
• License file is packaged (see here for an example).
• Source is from official source.
• Package does not vendor other packages. (If a package uses the source of another package, they should be separate packages or the licenses of all packages need to be packaged).
• If static libraries are linked in, the license of the static library is packaged.
• Package does not ship static libraries. If static libraries are needed, follow CFEP-18.
• Build number is 0.
• A tarball (url) rather than a repo (e.g. git_url) is used in your recipe (see here for more details).
• GitHub users listed in the maintainer section have posted a comment confirming they are willing to be listed there.
• When in trouble, please check our knowledge base documentation before pinging a team.

[hadim]

Another attempt to get deepspeed in conda-forge.

• Close Package request: deepspeed #18420

• Previous attempts:

  • Added recipe for DeepSpeed #14699
  • Add deepspeed #19021
  • Added recipe for deepspeed #19973

• This PR cannot be merged before ninja-python is available on conda-forge: Add python-ninja #19098

[conda-forge-webservices]

Hi! This is the friendly automated conda-forge-linting service.

I just wanted to let you know that I linted all conda-recipes in your PR (recipes/deepspeed) and found it was in an excellent condition.

[hadim]

@jaimergp @hmaarrfk: I have some questions regarding that recipe.

In short, pytorch is only needed at build time when the deepspeed ops are being built which can only happen when CUDA is available.

I tried to play with the selectors, so the CPU build is not dependent on a specific pytorch version. The goal is to reduce the number of unneeded builds here.

Please let me know what you think.

[hmaarrfk]

I guess you can build for slightly more cudas now, but the migrator will have a hard time updating you to the latest version of pytorch if you don't add the skips now.

This looks great!

[hadim]

Thanks @hmaarrfk. Regarding the skip for CPU and pytorch during build time, do you think this would not cause any issue for conda/mamba when it comes to pick the appropriate builds? I am afraid the CPU build will be picked more easily since it's not tight to a pytorch version versus the GPU builds that are tight to it (but maybe it's all good there?).

I guess you can build for slightly more cudas now

I guess I cannot do that in that PR and will wait for this to be merged.

[hmaarrfk]

GPU builds that are tight to it

It may prefer the GPU because it has fewer dependenices. Use a build number trick for most robustness
https://github.com/conda-forge/pytorch-cpu-feedstock/blob/main/recipe/meta.yaml#L4

[hmaarrfk]

Is there anything else you need from staged-recipies or can we merge?

[hadim]

I am ok to merge and address the ninja dep issue in the feedstock: #21708 (comment)

Also what should I do about CUDA arch above? Should I hard code the list?

[hmaarrfk]

Yes. please copy the list from pytorch.
https://github.com/conda-forge/pytorch-cpu-feedstock/blob/main/recipe/build_pytorch.sh#L94

[hadim]

@hmaarrfk LGTM for me here

[hmaarrfk]

  csrc/transformer/dropout_kernels.cu(102): error: no operator "*" matches these operands
              operand types are: __half2 * const __half2

  csrc/transformer/dropout_kernels.cu(103): error: no operator "*" matches these operands
              operand types are: __half2 * const __half2

  csrc/transformer/dropout_kernels.cu(216): error: no operator "*" matches these operands
              operand types are: __half2 * const __half2

  csrc/transformer/dropout_kernels.cu(217): error: no operator "*" matches these operands
              operand types are: __half2 * const __half2

  csrc/transformer/dropout_kernels.cu(335): error: no operator "*" matches these operands
              operand types are: __half2 * const __half2

  csrc/transformer/dropout_kernels.cu(336): error: no operator "*" matches these operands
              operand types are: __half2 * const __half2

you might have to raise the min version

[hadim]

@hmaarrfk seems ok now

[hadim]

Thanks @hmaarrfk !

[hmaarrfk]

no problem.

[h-vetinari]

Thanks for packaging this!

I think the following is not up to scratch though and should be fixed soon.