maxtext
maxtext copied to clipboard
Goodput integration
Integrate Goodput library with MaxText
This PR includes:
- Install Goodput dependency (ml-goodput-measurement in requirements.txt)
- Add config options to enable Goodput
- Update MaxText's train.py to use Goodput APIs to record and calculate Goodput
Tested: GCE
- Single Host v4-8 MaxText (example run: https://screenshot.googleplex.com/C7RsKp6TMHUHsn6)
- Multihost 2xv4-8 MaxText (example run: https://screenshot.googleplex.com/8ExSiZoN9oCaKq9)
- Multihost v4-16 MaxText (example run: https://screenshot.googleplex.com/46vdLJ9tVpCGxyZ)
GKE
- Single Host v4-8 MaxText (example run: https://screenshot.googleplex.com/4iDM9PBnAwfhDqZ)