Fixup third_party references

[Tobias-Fischer]

Checklist

• Used a personal fork of the feedstock to propose changes
• Bumped the build number (if the version is unchanged)
• Reset the build number to 0 (if the version changed)
• Re-rendered with the latest conda-smithy (Use the phrase @conda-forge-admin, please rerender in a comment in this PR for automated rerendering)
• Ensured the license file is being packaged.

[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 (recipe) and found it was in an excellent condition.

[Tobias-Fischer]

@conda-forge-admin, please rerender

[Tobias-Fischer]

With these changes, I can successfully compile the following test:

import torch
def fn(x):
    a = torch.cos(x).cuda()
    return a+1

input_tensor = torch.randn(10000).to(device="cuda:0")
new_fn = torch.compile(fn, backend="inductor")
a = new_fn(input_tensor)

print(a)

While it fails without the changes.

My big question mark is which package we need to add to the run dependencies to get the required ptxas and libdevice.10.bc- there have been some recent changes (see below) that I don't fully understand. To get my environment to work, I installed the cudatoolkit etc. from the nvidia channel.

See conda-forge/pytorch-cpu-feedstock#166
conda-forge/cudatoolkit-feedstock#62
conda-forge/staged-recipes#21382

/cc @jakirkham @h-vetinari @hmaarrfk

[h-vetinari]

This is great! :)

A nit about how the patches are checked in, for the run-requirements I'm hoping we can get some insights from @jakirkham / @leofang / @adibbley

[Tobias-Fischer]

It would be great if you could have another look @h-vetinari @hmaarrfk.

Also @jakirkham / @leofang / @adibbley it would be amazing if you could have a look to let us know which cuda packages are needed for ptxas.

[h-vetinari]

which cuda packages are needed for ptxas.

This was answered in the CUDA build order thread. It should be in the CUDA compiler package (though it might still get sliced up a bit differently)

[jakirkham]

which cuda packages are needed for ptxas.

This was answered in the CUDA build order thread. It should be in the CUDA compiler package (though it might still get sliced up a bit differently)

Indeed. xref: conda-forge/staged-recipes#21382 (comment)

A nit about how the patches are checked in, for the run-requirements I'm hoping we can get some insights from

Think I've missed the question here. Please let us know what you are looking for feedback on

[Tobias-Fischer]

Hi @h-vetinari - now that we pushed faiss over the line, do you think we can move forward here as well?

I think the only thing left here is that cudatoolkit on conda-forge does not contain the ptxas binary at the moment. However, the migration to the new cuda stuff will probably take quite some time (just guessing here). This PR only makes things better, and enables at least some users to use the new triton, whereas it's broken at the moment without the PR.

It would be great if we could merge here, and then open an issue that keeps track of the cuda migrations.

[h-vetinari]

Can you please confirm you're the author on all the patches you've added with your attribution, or point to the exact upstream PRs commits where they're taken from?

[h-vetinari]

OK, I'm fine to leave the cudatoolkit as a run-dep until we get the new infrastructure. I've also looked more closely and all patches seem to be coming from you (perhaps I misremembered from another PR where some were taken from unreleased upstream commits).

Basically, this LGTM. Any last comments @hmaarrfk?

[Tobias-Fischer]

Can you please confirm you're the author on all the patches you've added with your attribution, or point to the exact upstream PRs commits where they're taken from?

Confirmed :)

[h-vetinari]

LGTM, thanks a lot @Tobias-Fischer (I cleaned up the churn in the commit history a bit to avoid having to squash everything, hope that's OK; if not feel free to force-push your own cleaned up history)

@hmaarrfk, I'm planning to merge this in ~24h, unless of course you have other comments or need some more time to review here - please let us know!

[Tobias-Fischer]

LGTM, thanks a lot @Tobias-Fischer (I cleaned up the churn in the commit history a bit to avoid having to squash everything, hope that's OK; if not feel free to force-push your own cleaned up history)

Yes that's great, thanks a lot!

[Tobias-Fischer]

Kind ping @h-vetinari @hmaarrfk :)

[hmaarrfk]

Thank you all!

[h-vetinari]

Kind ping @h-vetinari @hmaarrfk :)

I knew I had wanted to merge something, but had forgotten which PR 😂
Sorry about the delay, and thanks for the PR!