`GradientCheckpointingLayer` checkpoints FSDP2 hooks; `use_reentrant=True` silently drops gradients when a layer runs more than once per bac

当 Transformers 的 GradientCheckpointingLayer 与 FSDP2 的 fully_shard 组合,且梯度检查点使用 use_reentrant=True 时,同一个 decoder layer 在单次 backward 前被多次 forward,该层梯度可能只




![[eager MOE] Selected expert IDs should probably not be treated as 0-d GPU tensors](https://www.chat-gpts.plus/wp-content/uploads/2026/11/49190-31d17efb-768x403.jpg)



![Muse tool parser JSON-decodes Optional[str] parameters, so "10001" becomes 10001](https://www.chat-gpts.plus/wp-content/uploads/2026/11/64546-39a741f7-768x403.jpg)