DeepSpeed
feat(zenflow): run the overlapped CPU optimizer in a native process
#8058
Merged

feat(zenflow): run the overlapped CPU optimizer in a native process #8058

delock merged 17 commits into master from zenflow-cpu-op
Antlera
Antlera Fix ZenFlow NaN under PyTorch-style backward via backward_prologue
d52a4ecb
Antlera Add fused multi-tensor CPU Adam for ZenFlow overlap step
1d9f3cce
Antlera Let CPU Adam kernel run serially without OpenMP
790a83a8
Antlera Add ZenFlowAdam: in-process overlapped CPU Adam
1640828d
Antlera Run ZenFlow stage 1/2 overlapped optimizer in-process
d60c7779
Antlera Run ZenFlow stage 1/2 overlapped optimizer in a separate native process
40491c90
Antlera Run ZenFlow stage 3 overlapped optimizer in the native process
4164f14d
Antlera Fail fast if the ZenFlow optimizer process does not start
60181d99
Antlera Stream ZenFlow optimizer copyback in chunks to bound the GPU memory peak
59daa9e1
Antlera Remove ZenFlow's superseded in-process overlapped optimizer path
2f9590bb
Antlera Recognize ZenFlowCPUAdam as a supported ZeRO optimizer
13ce892f
Antlera Antlera requested a review from tjruwase tjruwase 86 days ago
Antlera Antlera requested a review from tohtana tohtana 86 days ago
Antlera Antlera requested a review from loadams loadams 86 days ago
Antlera Antlera removed review request from tjruwase tjruwase 86 days ago
Antlera Antlera removed review request from tohtana tohtana 86 days ago
Antlera Antlera removed review request from loadams loadams 86 days ago
Antlera Antlera requested a review from delock delock 86 days ago
Antlera Antlera requested a review from tohtana tohtana 86 days ago
Antlera
chatgpt-codex-connector
chatgpt-codex-connector commented on 2026-06-10
Antlera Apply clang-format formatting to ZenFlowAdam definitions
1c582212
Antlera Fail loudly if the ZenFlow optimizer process dies mid-step
3a0d10ad
delock
delock commented on 2026-06-11
delock
delock commented on 2026-06-11
delock
delock commented on 2026-06-11
delock
delock commented on 2026-06-11
delock
delock commented on 2026-06-12
Antlera
delock
Antlera
Antlera Copy ZenFlow optimizer results back to the GPU without staging fp32
2b7f088f
Antlera Unit-test ZenFlow's zf/pt core-affinity split
b3b52df8
Antlera Remove the orphaned fused multi-tensor CPU Adam op
f935f589
Antlera Tidy ZenFlowAdam control block: drop redundant modulo, name the hp st…
667b8f23
Antlera
delock
delock approved these changes on 2026-07-07
delock delock merged 1ddbe484 into master 60 days ago
delock delock deleted the zenflow-cpu-op branch 60 days ago

Login to write a write a comment.

Login via GitHub

Assignees
No one assigned
Labels
Milestone