Lightning-AI / pytorch-lightning UNCLAIMED

Pretrain, finetune ANY AI model of ANY size on 1 or 10,000+ GPUs with zero code changes.

30975 0 0 Python

Fix: `no_grad` with AMP bug (#20921)

* Disable cache for torch.autocast in amp
* Add a test
* Only test for bf16-mixed
* Implement test to reproduce the issue

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Jirka Borovec <[email protected]>
(cherry picked from commit 216f9ec90c5bf3554f7cf484accee325f2a15440)

Bas Krahmer committed 7mo ago

2fe67a7724b96c3d5bd4bf938c6116d9011603b3

Parent: 120737c

Committed by Luca Antiga <[email protected]> on 9/5/2025, 1:14:02 PM