Lightning-AI / pytorch-lightning UNCLAIMED

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

0 0 104 Python

Synchronize gradients in manual optimization with DDPStrategy(static_graph=True) (#21251)

* fix: synchronize gradients in manual optimization with DDPStrategy(static_graph=True). Ensure gradients are reduced correctly when using manual optimization and DDP with static_graph enabled.

* Adds regression test to cover all combinations of optimization/static_graph.

* Initialize _pl_static_graph_delay_done attribute properly

* changelog

---------

Co-authored-by: Nicki Skafte Detlefsen <skaftenicki@gmail.com>

Sohaib Ahmed committed 10mo ago

10675b49d69b6ccd9ea0ee95484d180615b60ad6

Parent: 2f448e1

Committed by GitHub <noreply@github.com> on 10/29/2025, 5:39:26 AM