Accelerate's FSDP2 activation checkpointing was wrapping each child module instead of the layer, keeping every inter-child activation alive

GitHub·high signalXBlueskyLinkedInCopy link

accelerate 1.15.0, released 2026-09-09, fixes activation checkpointing wrapping `self_attn`, `mlp` and the norms individually rather than the matched transformer layer, so every activation between those children stayed saved for backward. PR #4172 wraps the layer itself. The release also adds an FSDP2-only `activation_checkpointing_offload` that moves per-layer checkpoint inputs to pinned CPU memory with gradients identical to plain checkpointing, and fixes FSDP2 + PEFT + `FULL_STATE_DICT` dropping every rank's adapter shard except rank 0 (#4206).