There was an autograd debugging line left in that significantly impacted performance. Discovered it while using a profiler to debug experiment speeds. Removing it reduced runtime for training my DAS benchmark work on pythia-14m from 289s to 170s!
Testing Done
No effect on library behaviours at all, just removed torch.autograd.set_detect_anomaly(True).
Checklist:
[x] My PR title strictly follows the format: [Your Priority] Your Title
Description
There was an autograd debugging line left in that significantly impacted performance. Discovered it while using a profiler to debug experiment speeds. Removing it reduced runtime for training my DAS benchmark work on
pythia-14m
from 289s to 170s!Testing Done
No effect on library behaviours at all, just removed
torch.autograd.set_detect_anomaly(True)
.Checklist:
[Your Priority] Your Title