Pull Requests microsoft/DeepSpeed

fix: remove premature MPI environment variable check in OpenMPIRunner

#7751 by nathon-lee was merged 2026-01-01 08:33

feat: add parameter-level precision control for BF16 training

#7750 by nathon-lee was closed 2026-03-02 03:20

Removed amp testcases

#7745 by k-artem was merged 2026-01-05 15:53

Fix multiprocessing testcase

#7743 by k-artem was merged 2026-01-15 11:40

Fix DecoupledCheckpointEngine deadlock and improve reliability

#7742 by Rakshit-gen was merged 2025-12-22 15:25

Fix Nebula checkpoint engine commit() API mismatch

#7740 by Rakshit-gen was merged 2025-12-22 13:07

fix(issue-7701): un-ignore .cuh under deepspeed/ops so multi_tensor_a…

#7739 by nathon-lee was closed 2026-01-07 08:52

Add core api update blog

#7738 by tohtana was merged 2025-12-19 20:57

[BUG] Fix UlyssesSPAttentionHF.register_with_transformers() crash with PEFT models

#7737 by Rakshit-gen was merged 2025-12-19 18:57

Fix OnebitLamb NaN propagation with empty parameters

#7736 by Rakshit-gen was merged 2025-12-24 18:49

Fix #7733: Replace torch.sqrt with math.sqrt in scale_lr for sqrt method

#7735 by Rakshit-gen was merged 2025-12-19 15:28

replace moe checkpoint dp_world_size with seq_dp_world_size

#7732 by wukong1992 was merged 2025-12-19 18:33

Fix testcases that depends on triton

#7731 by k-artem was merged 2025-12-17 15:56

Fix rare hang in DeepSpeed Async I/O wait by releasing the Python GIL

#7727 by xylian86 was merged 2025-12-18 18:21

Skip none in backward hook

#7725 by tohtana was merged 2025-12-12 06:39

[Engine] Only scale gradients if scale_wrt_gas is True

#7724 by kashif was merged 2025-12-12 10:34

fix typo

#7722 by stas00 was merged 2025-12-12 00:11

Disable deterministic option in compile tests

#7720 by tohtana was merged 2025-12-09 23:52

Update version

#7719 by sfc-gh-truwase was merged 2025-12-09 14:59

fix: The different lr for Muon doesn't work

#7716 by KeeProMise was closed 2025-12-09 08:32

Fix SuperOffloadOptimizer_Stage3 crash due to missing param_names parameter

#7715 by ImaGoodFella was merged 2025-12-10 00:18

Wall clock timers API

#7714 by sfc-gh-truwase was merged 2025-12-09 07:18

fix: skip aio wait when swap tensors is empty

#7712 by xylian86 was merged 2025-12-04 03:25

Fix that ds_secondary_tensor may be dirty when loading the model or zero checkpoint for zero++.

#7707 by zhengchenyu was merged 2025-12-03 21:46

Add news about Ray x DeepSpeed Meetup

#7704 by PKUWZP was merged 2025-11-24 15:26

Low-precision master params/grads/optimizer states

#7700 by tohtana was merged 2025-12-04 03:53

Expand FlopsProfiler stale counter documentation

#7699 by cklxx was closed 2025-11-30 12:49

Trust intel server for XPU tests

#7698 by tohtana was merged 2025-11-18 23:42

Add Qwen2.5 to AutoTP model list

#7696 by delock was merged 2025-11-18 16:06

Update SECURITY.md to point to GitHub reporting rather than Microsoft

#7692 by loadams was merged 2025-11-18 16:06