GHA that adds Flash Attention Benchmarking on B200 #49
Conversation
jduprat
changed the title
jduprat
force-pushed
the
flash-attention-benchmarking
branch
2 times, most recently
from
e2a6508 to
4b6416c
Compare
| push: | ||
| paths: | ||
| - .github/workflows/flash_attention.yml | ||
| repository_dispatch: |
There was a problem hiding this comment.
Oh, this is new to me. What does this repository_dispatch do? :)
There was a problem hiding this comment.
This allows triggering the action by running —
$ curl -L -X POST -H "Accept: application/vnd.github+json" -H "X-GitHub-Api-Version: 2022-11-28" -H "Authorization: Bearer $TOKEN" https://api.github.com/repos/pytorch/pytorch-integration-testing/dispatches -d '{"event_type": "benchmark_flash_attention"}'
The intent is trigger this action when there is a commit in the FA repo, but the FA repo needs to be updated.
jduprat
force-pushed
the
flash-attention-benchmarking
branch
from
4b6416c to
7536545
Compare
Checkout the FA repo and run benchmark as part of this action
jduprat
force-pushed
the
flash-attention-benchmarking
branch
from
7536545 to
3b8c5bd
Compare
| echo '<h1>B200 1000W</h1>' >> $GITHUB_STEP_SUMMARY | ||
| nvidia-smi | ||
| export PYTHONPATH=$(pwd) | ||
| python benchmarks/benchmark_attn.py >> $GITHUB_STEP_SUMMARY |
There was a problem hiding this comment.
If you are interested in building a dashboard out of it on https://hud.pytorch.org later, feel free to take a look at https://github.com/pytorch/pytorch/wiki/PyTorch-OSS-benchmark-infra#benchmark-results-format to see if we could save the attention benchmark output here in the same format. That's all it is needed to be done.
Triggered on a schedule (every 2 hours for now), by hand, or by remote trigger.
Checkout the FA repo and run its benchmark