Accelerate 1.15.0 fixes FSDP2 checkpointing that leaked memory

Hugging Face released Accelerate 1.15.0 on September 9, fixing a bug in FSDP2's activation checkpointing.

The bug kept memory allocated that checkpointing exists to free. That cuts into the savings teams rely on when training large models across multiple GPUs.

The wrapping logic checkpointed self_attn, mlp and the norm layers on their own instead of the matched transformer layer that contains them. Every activation passed between those child modules stayed alive through the backward pass rather than being freed and recomputed. PR #4172 moves the wrap up to the full layer.

The same Accelerate 1.15.0 release notes list a second fix that's worse in a quieter way. Combining FSDP2, PEFT and FULL_STATE_DICT dropped every rank's adapter shard except rank 0's when saving a checkpoint.

The file writes without error, it loads without error, and nothing in that path tells you weights are missing. Anyone running that combination ends up training or serving with a fraction of the adapter they saved.