DeepSpeed
Restore fp32 masters for ZeRO-3 and the ZeRO-1 BF16 optimizer
#8892
Open

Restore fp32 masters for ZeRO-3 and the ZeRO-1 BF16 optimizer #8892

sfc-gh-truwase
sfc-gh-truwase Keep fp32 masters on ZeRO-1 and ZeRO-2 when a resume skips optimizer …
f9958d4d
sfc-gh-truwase Restore fp32 masters for ZeRO-3 and the ZeRO-1 BF16 optimizer.
bb31f403
sfc-gh-truwase sfc-gh-truwase requested a review from tjruwase tjruwase 4 days ago
sfc-gh-truwase sfc-gh-truwase requested a review from tohtana tohtana 4 days ago
sfc-gh-truwase sfc-gh-truwase requested a review from loadams loadams 4 days ago
chatgpt-codex-connector
chatgpt-codex-connector
chatgpt-codex-connector commented on 2026-10-22
sfc-gh-truwase Merge master into the ZeRO-1/2 fp32 master branch.
b3728af3
sfc-gh-truwase Clarify FP32 checkpoint loading paths
e489f80b
sfc-gh-truwase Keep saved fp32 masters on every full ZeRO load.
a7ce02d9
sfc-gh-truwase Fail a universal load that is missing its fp32 master.
2f56cafc
sfc-gh-truwase Skip DistributedFixture tests when the world size is unavailable.
894a2b72
sfc-gh-truwase Merge the ZeRO-1/2 fp32 master branch into the ZeRO-3 and BF16 follow…
e3b65b61
sfc-gh-truwase Test ZeRO-3 and BF16 fp32 masters when optimizer state is skipped.
b190a3f7

Login to write a write a comment.

Login via GitHub

Assignees
No one assigned
Labels
Milestone