Add lr scheduler, weight decay and max_grad_norm#214

Merged

orangetin merged 12 commits intomainfrom arsh/lr-scheduler

Nov 19, 2024

+95-4

Commits on Nov 14, 2024

Add weight decay and max_grad_norm

Arsh Zahed
authored and
Arsh Zahed
committed

Commits on Nov 15, 2024

Change to min_lr_ratio

Arsh Zahed
authored and
Arsh Zahed
committed

Commits on Nov 16, 2024

Update max_grad_norm and weight_decay defaults, supported vals

Arsh Zahed
authored and
Arsh Zahed
committed
Add hints for disabling max_grad_nomr

Arsh Zahed
authored and
Arsh Zahed
committed
add back learning rate hint

Arsh Zahed
authored and
Arsh Zahed
committed
add max_grad_norm and weight_decay to FinetuneRequest

Arsh Zahed
authored and
Arsh Zahed
committed
Remove percentage from min_lr_ratio description

Arsh Zahed
authored and
Arsh Zahed
committed
Use training_args

Arsh Zahed
authored and
Arsh Zahed
committed

Commits on Nov 18, 2024

Fix hints and typing

Arsh Zahed
authored and
Arsh Zahed
committed
Update version to 1.3.5

Arsh Zahed
authored and
Arsh Zahed
committed
Fix more typing

Arsh Zahed
authored and
Arsh Zahed
committed

Commits on Nov 19, 2024

Make min_lr_ratio optional

Arsh Zahed
authored and
Arsh Zahed
committed