flash-attn v2.6.3 + TORCH_CUDA_ARCH_LIST=8.0;8.6;8.9;9.0+PTX

[regro-cf-autotick-bot]

It is very likely that the current package version for this feedstock is out of date.

Checklist before merging this PR:

• Dependencies have been updated if changed: see upstream
• Tests have passed
• Updated license if changed and license_file is packaged

Information about this PR:

1. Feel free to push to the bot's branch to update this PR if needed.
2. The bot will almost always only open one PR per version.
3. The bot will stop issuing PRs if more than 3 version bump PRs generated by the bot are open. If you don't want to package a particular version please close the PR.
4. If you want these PRs to be merged automatically, make an issue with @conda-forge-admin,please add bot automerge in the title and merge the resulting PR. This command will add our bot automerge feature to your feedstock.
5. 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.

Pending Dependency Version Updates

Here is a list of all the pending dependency version updates for this repo. Please double check all dependencies before merging.

Name	Upstream Version	Current Version
pytorch-cpu	131547

Dependency Analysis

We couldn't run dependency analysis due to an internal error in the bot, depfinder, or grayskull. :/ Help is very welcome!

_{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 - please use this URL for debugging.}

Closes #11
Closes #17

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

[carterbox]

Looks like we are exceeding the 6 hour limit even when building for only a single arch. Looks like we need to finish the process to get on the Quantsight GPU server.

@weiji14, could you please agree to the terms of use for the Quantsight GPU severs by opening a pull request here: https://github.com/Quansight/open-gpu-server

Then we can open a request to have the bot enable the Quantsight GPU server for this feedstock using this template: https://github.com/conda-forge/admin-requests/blob/main/examples/example-open-gpu-server.yml

Ref: https://conda-forge.org/docs/maintainer/knowledge_base/#packages-that-require-a-gpu-or-long-running-builds

[weiji14]

Sure, let me work on that later, recent builds have been too close to the 6hr mark and I've had to restart CI for the past two releases already, and would be good to handle more CUDA archs (xref #4). Not sure if we actually need a GPU-enabled server though, maybe just a longer timeout (+more CPU/RAM)?

[carterbox]

Not sure if we actually need a GPU-enabled server though, maybe just a longer timeout (+more CPU/RAM)?

The only way to accomplish builds longer than 6 hours is to use Quantsight's servers. I'm not sure what options Quantsight provides for servers with/without GPUs, but this package is a CUDA enabled package, so we should just use the GPU server list.

[carterbox]

TODO:

• Enable other build archs xref: Build on OpenStack server for TORCH_CUDA_ARCH_LIST=8.0;8.6;8.9;9.0+PTX #4

[carterbox]

On the openstack runners, the jobs are crashing during the first compilation phase. Trying to reduce the number of concurrent compilations to 2 in order to prevent this crash (from a max of 4).

[carterbox]

Maybe they're not crashing? Maybe the jobs are just being killed at 30mins?

[carterbox]

@conda-forge-admin, please rerender

[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.

[weiji14]

Thanks @jakirkham for the merge tips 🙏

The previous test build with CUDA 11.8 / Python 3.12 seems to have completed successfully on cirun-openstack-cpu-xlarge (32GB RAM) with MAX_JOBS=4 at commit 5878e8b in 8h6m. See logs at https://github.com/conda-forge/flash-attn-feedstock/actions/runs/11246603140/job/31367673505

I'm now testing CUDA 12.0 / Python 3.13 (cherry-picking from #17) at commit e4e877b, and if this builds complete within the time limit, I think we should have enough info on what resources will be needed for the full matrix builds.

[weiji14]

Ok, are we ready for the full matrix builds?

[carterbox]

Sounds good to me! 8 hours per build seems reasonable.

[carterbox]

@conda-forge-admin, please rerender

[carterbox]

Wait. Setuptools is not a runtime dependency of this package. Pytorch needs to fix its missing/extra dependency, not us.

[carterbox]

@conda-forge-admin, please rerender

[carterbox]

@conda-forge-admin, please rerender

[jakirkham]

Thanks Daniel! 🙏

Looks like cirun doesn't like the bot commit. So may need to push another commit

[weiji14]

Ok, re-rendered manually using conda smithy rerender -c auto. I've set the timeout back to 9 hours.

[weiji14]

Builds all pass now 🎉 Longest one was CUDA 11.8 / Python 3.12 that took 8h 26m 41s (under the 9h time limit).

Shall we wait for conda-forge/pytorch-cpu-feedstock#276, and then we can include the Python 3.13 builds here too?

[weiji14]

Remove after conda-forge/pytorch-cpu-feedstock#276 is merged (remember to re-render manually afterwards).

Suggested change

skip: true # [py==313] # Skip until pytorch dependency on setuptools is fixed

[carterbox]

We have to rebuild to add those additional modules requested in #18. We can see if pytorch is ready for 3.13 then.

[jakirkham]

Hooray! 🥳

Excellent work 😄 Thank you both! 🙏