PR #48640 Fix deepspeed ci

Fix deepspeed ci #48640

ydshieh merged 4 commits into main from fix-deepspeed-ci

SunMarc force pushed from 8f093f28 to 175805d5 11 days ago

SunMarc force pushed from 175805d5 to 237ec57e 11 days ago

SunMarc force pushed from 237ec57e to 307f043d 11 days ago

SunMarc force pushed from 307f043d to 5a842f8c 11 days ago

SunMarc force pushed from 5a842f8c to 4a09f792 11 days ago

SunMarc force pushed from 4a09f792 to a554efc0 11 days ago

SunMarc force pushed from a554efc0 to ebe017c1 11 days ago

Rebuild the DeepSpeed CI image on a plain CUDA 12.6 base

4a7ea01a

Fix DeepSpeed Trainer tests against deepspeed>=0.19.6

7de47902

Give the grad accumulation grad-norm check its own tolerance

3e84c512

SunMarc force pushed from ebe017c1 to 3e84c512 11 days ago

SunMarc requested a review from

ydshieh 11 days ago

Give the distributed test collator position_ids for Ulysses SP

ecc4b740

SunMarc force pushed from 101f7dbf to ecc4b740 11 days ago

sergereview commented on 2026-09-09

ydshieh approved these changes on 2026-09-09

ydshieh merged 29723434 into main 10 days ago

ydshieh deleted the fix-deepspeed-ci branch 10 days ago

Reviewers

ydshieh

sergereview

Assignees

No one assigned

Labels

None yet

Milestone

No milestone