Rebuild for pytorch 2.5, add python 3.13, update to flash-attn 2.7.0.post2

[regro-cf-autotick-bot]

This PR has been triggered in an effort to update pytorch25.

Notes and instructions for merging this PR:

1. Please merge the PR only after the tests have passed.
2. Feel free to push to the bot's branch to update this PR if needed.

Please note that if you close this PR we presume that the feedstock has been rebuilt, so if you are going to perform the rebuild yourself don't close this PR until the your rebuild has been merged.

---

If this PR was opened in error or needs to be updated please add the bot-rerun label to this PR. The bot will close this PR and schedule another one. If you do not have permissions to add this label, you can use the phrase @conda-forge-admin, please rerun bot in a PR comment to have the conda-forge-admin add it for you.

_{This PR was created by the regro-cf-autotick-bot. The regro-cf-autotick-bot is a service to automatically track the dependency graph, migrate packages, and propose package version updates for conda-forge. Feel free to drop us a line if there are any issues! This PR was generated by https://github.com/regro/cf-scripts/actions/runs/11669226258 - please use this URL for debugging.}

Closes #22
Closes #17

[conda-forge-admin]

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/meta.yaml) and found it was in an excellent condition.

[jeongseok-meta]

@weiji14 I was able to build for linux with no errors (https://anaconda.org/jslee02/flash-attn/files). Could you please retrigger the CI to see if this build?

[weiji14]

Sure, I can restart the CI, but didn't want to use up most of the Openstack CI runners if you're still working on #19 (they have a limited amount of runners, and these flash-attn builds take ~9 hours as you're aware). Would it be possible to finish #19 first and then we do Pytorch 2.5 rebuild + Python 3.13 (#17)?

[jeongseok-meta]

I wasn't aware of #19 and the order of preference for which one should be merged. It's totally up to you and the author of #19, but I assumed this one would be a simple merge (if it passes) while the other, #19, seems to need more work to reach consensus.

[weiji14]

Oh sorry, I was confusing you with the author of #19 (been looking at too many feedstocks recently 😅). This feedstock is a little different to the other conda-forge feedstocks in that 1) We're using a separate Openstack CI system that has a limited amount of resources compared to Azure Pipelines, and 2) The CI takes almost 9 hours to build per combination, and with 5 Python versions * 2 CUDA versions = 10 runs, this can block up the server queue (as mentioned before at #10 (comment)).

So I think we'll wait for #19 to finish first, will be easier to rebase this one (#20) on top of that later.

[jeongseok-meta]

No worries! The plan makes sense to me, and it sounds good. In the meantime, I've pushed my local build for pytorch 2.5, so if someone wants to try flash-attn with pytorch 2.5, they can give it a shot!

https://anaconda.org/jslee02/flash-attn/files

[carterbox]

We will merge the other pending branches into this branch for the next release.

[carterbox]

Waiting for #19 to complete in order to publish rebase because I don't want to monopolize the build resources. I already have it rebased locally.

[carterbox]

No automerge because the bot doesn't have permission to trigger builds. Anyone please merge if everything passes.

[weiji14]

There was a download error because the 2.7.0 release wasn't on PyPI, so I've bumped it to 2.7.0.post2 (https://pypi.org/project/flash-attn/2.7.0.post2/).

I've also checked the diff on setup.py between 2.6.3 and 2.7.0.post2, seems like only ROCm related changes which shouldn't affect what's here in this feedstock (which is CUDA-only).

[weiji14]

All builds passed except for Python 3.13 + CUDA 11.8 which seemed to have hit a timeout error:

conda.CondaMultiError: ('Connection broken: IncompleteRead(282831187 bytes read, 206660643 more expected)', IncompleteRead(282831187 bytes read, 206660643 more expected))
('Connection broken: IncompleteRead(282831187 bytes read, 206660643 more expected)', IncompleteRead(282831187 bytes read, 206660643 more expected))

Do we want to restart that build to make sure it works? I've tested building locally up to the [1/85] step, but unsure if it'll get to [85/85] 🙂

[weiji14]

I'll just approve and merge, assuming that the timeout at #20 (comment) is harmless 🤞

[carterbox]

We don't need to use .post in our version number because the changes post releases don't affect our builds. They are just a way for upstream to rebuild their wheels because PYPI doesn't support build numbers.

[weiji14]

Yeah, but we're pulling from PyPI, and 2.7.0 wasn't published to PyPI (see https://pypi.org/project/flash-attn/2.7.0.post2/#history). We would have needed to hardcode url: https://pypi.org/packages/source/f/flash-attn/flash_attn-{{ version }}.tar.gz to include the .post2 suffix here if keeping to the 2.7.0 version number:

flash-attn-feedstock/recipe/meta.yaml

Lines 1 to 8 in 82abb88

	{% set version = "2.7.0.post2" %}
	package:
	name: flash-attn-split
	version: {{ version }}
	source:
	- url: https://pypi.org/packages/source/f/flash-attn/flash_attn-{{ version }}.tar.gz