Support AutoEP with ZeRO-3 zero.Init source modules #8060
tohtana
merged
21 commits into
deepspeedai:master
from
tohtana:tohtana/autoep-zero3-zero-init
Enable AutoEP ZeRO-3 placement groups
105f0e8c
Fix ZeRO-3 multi-gather wait forwarding
4232ea36
Gather AutoEP source ZeRO params
807085c8
Document constrained AutoEP ZeRO-3 support
28e725ab
Fix AutoEP zero.Init unit test import
5770e0f9
Add 8 GPU AutoEP zero.Init validation
573b5251
Fix AutoEP ZeRO-3 expert gradient averaging
8c05ea76
Add partition-native AutoEP ZeRO-3 checkpoints
d7fac4a9
Address review findings for partition-native AutoEP checkpoints
8f528ced
Fail fast on pre-partitioned AutoEP expert params with mismatched groups
b76fb577
Update AutoEP checkpoint docs for ZeRO-3 partition-native support
c177dbce
Validate AutoEP universal topology loads
fd4a861c
Fix AutoEP topology fixture registration
bac569d8
Register AutoEP topology baseline as an explicit pytest fixture
6e7400d2
Fix AutoEP topology universal router check
fbb1bf91
Update config-json AutoEP ZeRO-3 universal topology-change note
18569fc0
Support AutoEP universal module-only loads
50f65163
Preserve fp32 masters for AutoEP module-only loads
f2d422d4
delock
commented
on 2026-06-16
delock
commented
on 2026-06-16
delock
commented
on 2026-06-16
delock
commented
on 2026-06-17
Address AutoEP ZeRO-3 review comments
a90deece
Merge upstream master into AutoEP ZeRO-3 PR
dbcd098a
delock
approved these changes
on 2026-06-25
Merge branch 'master' into tohtana/autoep-zero3-zero-init
dca6e1af
Assignees
No one assigned
Labels
None yet
Milestone
No milestone
Login to write a write a comment.
Login via GitHub