pytorch / opacus

Training PyTorch models with differential privacy
https://opacus.ai
Apache License 2.0
1.67k stars 332 forks source link

relax benchmarking thresholds for new PyTorch #533

Closed ffuuugor closed 1 year ago

ffuuugor commented 1 year ago

Since PyTorch 1.13 was released, our runtime/memory benchmarks are failing on the recurrent modules (GRU/RNN/LSTM)

The reason is - our implementation did not change, but the pytorch baseline got faster - by roughly the factor of 3.

Changing the thresholds based on the latest performance on CircleCI machine (+15% margin)

facebook-github-bot commented 1 year ago

@ffuuugor has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.