pytorch

ebd4c165 - Back out "`GradScaler` recomputes `optimizer_state["found_inf_per_device"]` before `optimizer.step` (#97415)" (#98613)

Back out "`GradScaler` recomputes `optimizer_state["found_inf_per_device"]` before `optimizer.step` (#97415)" (#98613) Summary: This change causes multi-GPU job from XI team to hang after 8K steps. Differential Revision: D44797248 Pull Request resolved: https://github.com/pytorch/pytorch/pull/98613 Approved by: https://github.com/ngimel