[WIP]optimize memory usage of fsdp2 init model#117

Closed

kevssim wants to merge 29 commits intomodelscope:mainfrom

kevssim:optimize_fsdp_init

Commits on Mar 18, 2026

feat(fsdp2): add _broadcast_sharded_state_dict, _get_non_persistent_buffers, _restore_non_persistent_buffers helpers
kevssim
committed
feat(fsdp2): enable cpu_ram_efficient_loading for both strategies; pass memory_efficient through _lazy_wrap_model
kevssim
committed
refactor(fsdp2): use _non_persistent_buffers_set for precise non-persistent buffer detection
kevssim
committed
Merge branch 'optimize_fsdp_init' of https://github.com/kevssim/twinkle into optimize_fsdp_init
kevssim
committed
wip
kevssim
committed
test(fsdp2): make tests platform-agnostic (cuda/npu) via Platform API
kevssim
committed
fix(test): pass inputs as List[InputFeature] to forward_backward
kevssim
committed
fix(test): add position_ids to e2e test batch
kevssim
committed
fix(test): simplify e2e test to only verify wrap_model, avoid processor device issues
kevssim
committed

Commits on Mar 19, 2026