DeepSpeed
[AffineIR] Affine storage maps for live ZeRO parameters and optimizer states
#8885
Merged

[AffineIR] Affine storage maps for live ZeRO parameters and optimizer states #8885

jinyouzhi
jinyouzhi Add affine storage maps for ZeRO parameters and optimizer states
95934d8c
jinyouzhi Compose layout maps with their storage maps
d82532e4
jinyouzhi Describe AutoTP parameter storage under ZeRO
f5f9015c
jinyouzhi Specify live storage maps for phase 2
f32a995e
jinyouzhi Describe nested storage with N-D boxes
17143322
jinyouzhi Gather optimizer fragment ranges in a gloo-compatible shape
ca87f28b
jinyouzhi Fix stat size lookup in checkpoint conversion test
683217d2
jinyouzhi jinyouzhi changed the title Affine storage maps for live ZeRO / AutoTP parameters and optimizer states [AffineIR] Affine storage maps for live ZeRO / AutoTP parameters and optimizer states 7 days ago
jinyouzhi jinyouzhi changed the title [AffineIR] Affine storage maps for live ZeRO / AutoTP parameters and optimizer states [AffineIR] Affine storage maps for live ZeRO parameters and optimizer states 7 days ago
jinyouzhi Merge branch 'master' into zero-storage-affine-map
33125856
jinyouzhi Keep nested boxes dense in every replica and support scalars
2d8d873e
jinyouzhi jinyouzhi marked this pull request as ready for review 7 days ago
jinyouzhi jinyouzhi requested a review from tjruwase tjruwase 7 days ago
jinyouzhi jinyouzhi requested a review from loadams loadams 7 days ago
jinyouzhi jinyouzhi requested a review from tohtana tohtana 7 days ago
chatgpt-codex-connector
chatgpt-codex-connector
chatgpt-codex-connector commented on 2026-10-22
jinyouzhi Handle zero.Init AutoTP, gathered and NVMe ZeRO-3 parameter storage
e34367a7
delock delock requested a review from delock delock 7 days ago
jinyouzhi Merge branch 'master' of https://github.com/deepspeedai/DeepSpeed int…
42bafc57
0z5a
jinyouzhi Test flat storage on the write side and gather shards without their b…
7f927940
jinyouzhi
jinyouzhi
jinyouzhi
Achyuthan-S
0z5a
delock
delock approved these changes on 2026-10-28
delock delock merged f8185049 into master 1 day ago

Login to write a write a comment.

Login via GitHub

Assignees
No one assigned
Labels
Milestone