[FSDP2] cpu_ram_efficient_loading materializes full model on CPU for every non-rank-0, defeating rank-0 broadcast (host OOM)
![[FSDP2] cpu_ram_efficient_loading materializes full model on CPU for every non-rank-0, defeating rank-0 broadcast (host OOM)](https://www.chat-gpts.plus/wp-content/uploads/2026/11/49297-32f2ed2d-768x403.jpg)
这个报错通常出现在多卡、单机多进程的 FSDP2 训练场景中:用户启用了 cpu_ram_efficient_loading ( FSDP_CPU_RAM_EFFICIENT_LOADING=1 ),本以为只有 rank-0 在 CPU 上加载权重、再广播给其他 rank,但 from_pretra




![[Bug]: GLM-5.x (DSA) under pipeline parallelism: reuse layers at a stage start read stale top-k indices](https://www.chat-gpts.plus/wp-content/uploads/2026/11/63170-1a708cbf-768x403.jpg)
![[bug]: Qwen-Image Edit reference-image tokens get 1-D positions under transformers 5](https://www.chat-gpts.plus/wp-content/uploads/2026/11/9983-684faac6-768x403.jpg)


