Support TP-compatible Torch AMP and Update trainer API (#27)

* Add gradient accumulation, fix lr scheduler

* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)

* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes

* fixed trainer

* Revert "fixed trainer"

This reverts commit 2e0b0b7699.

* improved consistency between trainer, engine and schedule (#23)

Co-authored-by: 1SAA <c2h214748@gmail.com>

Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
This commit is contained in:
Frank Lee
2021-11-18 19:45:06 +08:00
committed by GitHub
parent 2b05de4c64
commit 3defa32aee
80 changed files with 2194 additions and 1584 deletions

View File

@@ -66,11 +66,10 @@ class CosineAnnealingWarmupLR(WarmupScheduler):
:type last_epoch: int, optional
"""
def __init__(self, optimizer, total_steps: int, warmup_steps: int = 0, eta_min: int = 0, last_epoch: int = -1,
**kwargs):
def __init__(self, optimizer, total_steps: int, warmup_steps: int = 0, eta_min: int = 0, last_epoch: int = -1):
base_scheduler = _CosineAnnealingLR(
optimizer, total_steps - warmup_steps, eta_min=eta_min)
super().__init__(optimizer, warmup_steps, base_scheduler, last_epoch=last_epoch)
optimizer, total_steps - warmup_steps, eta_min=eta_min, last_epoch=last_epoch)
super().__init__(optimizer, warmup_steps, base_scheduler)
@LR_SCHEDULERS.register_module