DeepSpeed
Add AutoEP + AutoTP parallel folding
#8064
Open

Add AutoEP + AutoTP parallel folding #8064

tohtana
tohtana Add AutoEP + AutoTP parallel folding
278c9194
tohtana tohtana requested a review from tjruwase tjruwase 23 days ago
tohtana tohtana requested a review from hwchen2017 hwchen2017 23 days ago
tohtana tohtana requested a review from loadams loadams 23 days ago
tohtana tohtana requested a review from GuanhuaWang GuanhuaWang 23 days ago
chatgpt-codex-connector
chatgpt-codex-connector commented on 2026-06-13
tohtana Fix folded TP gradient reductions
0d44a1d2
tohtana Normalize folded TP ZeRO gradients
8b1c0428
PKUWZP PKUWZP requested a review from PKUWZP PKUWZP 22 days ago
delock
delock commented on 2026-06-18
delock
delock commented on 2026-06-19
tohtana Fix AutoEP folded gradient strategy
7246b146
tohtana Document AutoEP folding gradient-reduction rationale
e5c1ba23
tohtana Generalize AutoEP+AutoTP folding to cross-lane expert parallelism
13dceecf
PKUWZP
tohtana Gate AutoEP folding routing validation
59fdca46
tohtana Apply AutoEP folding yapf formatting
7426b121
tohtana Merge upstream master into AutoEP folding branch
ef57cb43
delock
delock approved these changes on 2026-07-01
PKUWZP
PKUWZP requested changes on 2026-07-01
tohtana Merge upstream master into AutoEP folding branch
f2c7f731
tohtana Add folded DeepCompile rejection tests
e6239e4a
tohtana Reject folded AutoEP DeepCompile configs
3c1e91a8
tohtana Add BF16 folded HP grad correction test
3fcc33bd
tohtana Apply folded correction before BF16 HP grad accumulation
b485a75e
tohtana Add ZeRO-1 overlap folding correction tests
4906a3a6
tohtana Apply folded correction in ZeRO-1 overlap hooks
42c4148d
tohtana Handle DeviceMesh split fallback for AutoTP groups
96f3b3b7
tohtana Apply pre-commit formatting
bc28dd8a

Login to write a write a comment.

Login via GitHub

Assignees
No one assigned
Labels
Milestone