DeepSpeed
Route ZeRO/SuperOffload pin sites through accelerator pin_memory
#8256
Open

Route ZeRO/SuperOffload pin sites through accelerator pin_memory #8256

sfc-gh-truwase wants to merge 7 commits into master from tjruwase/pin-memory-route-zero
sfc-gh-truwase
sfc-gh-truwase Route ZeRO/SuperOffload pin sites through accelerator pin_memory
37a5bf18
sfc-gh-truwase sfc-gh-truwase requested a review from tjruwase tjruwase 9 days ago
sfc-gh-truwase sfc-gh-truwase requested a review from tohtana tohtana 9 days ago
sfc-gh-truwase sfc-gh-truwase requested a review from loadams loadams 9 days ago
chatgpt-codex-connector
sfc-gh-truwase Fix CPU CI skip for offload pin-route unit test
e6dc5f9d
sfc-gh-truwase sfc-gh-truwase requested a review from delock delock 6 days ago
sfc-gh-truwase sfc-gh-truwase requested a review from xylian86 xylian86 6 days ago
delock
delock approved these changes on 2026-08-20
delock
sfc-gh-truwase
xylian86
xylian86 commented on 2026-08-20
sfc-gh-truwase Keep pinned reload sources alive until ZeRO-1/2 synchronizes.
aebb2819
xylian86
xylian86 commented on 2026-08-24
sfc-gh-truwase Allocate pinned memory directly when pin_memory is not copying.
246c02e1
delock Merge branch 'master' into tjruwase/pin-memory-route-zero
598ad17a
xylian86
xylian86 approved these changes on 2026-08-24
tohtana
tohtana requested changes on 2026-08-24
sfc-gh-truwase Release offload_states pinned buffers in ZeRO destroy().
05ddc1a4
sfc-gh-truwase Honor offload_optimizer.pin_memory for the SuperOffload worker buffer.
658dc247

Login to write a write a comment.

Login via GitHub

Assignees
No one assigned
Labels
Milestone