microsoft/DeepSpeed

Pull Requests Commits

ci: restore Modal parity for focused test selection

tohtana committed 7 days ago

Verified 080d13c5

ci: install diagnostic source pins separately

tohtana committed 7 days ago

Verified 7f72d8b3

ci: allow pinned prerelease diagnostics

tohtana committed 7 days ago

Verified 8685331c

ci: add bounded Modal distributed reproducer

tohtana committed 7 days ago

Verified 6838418f

Merge branch 'master' into yh0903/bf16-release-lp-grads

tohtana committed 8 days ago

Verified 34704e11

feat(rollout): align decode fronts and trim dead zones (#8497)

nathon-lee committed 9 days ago

Verified a77aeb67

Add an opt-in fused Triton RoPE for Hugging Face split-half models (#8663)

yh0903 committed 9 days ago

Verified 583fc9c3

Merge branch 'master' into yh0903/bf16-release-lp-grads

tohtana committed 9 days ago

Verified ede9a9f2

Give rollout cache fakes the per-layer config Transformers now reads (#8730)

sfc-gh-truwase committed 10 days ago

Verified 8fb2c9a2

Harden ZeRO-1/2 offload gradient storage lifetime and stream ordering (#8632)

jeffra committed 11 days ago

Verified c687df79

Fix Ulysses SP registration for string model paths and gloo collectives (#8702)

delock committed 11 days ago

Verified 04e0be2b

AutoEP: reorder expert rows with gathers on the collective path (#8671)

yh0903 committed 11 days ago

Verified bfd29175

AutoEP: MiniMax-M3 preset (#8651)

pengdurice committed 11 days ago

Verified b9ce5f42

Add opt-in fused AutoEP DeepEP row weighting (#8644)

yh0903 committed 11 days ago

Verified 2c4ead4b

Add Fengchun Hua (hipudding) to COMMITTERS.md (#8718)

PKUWZP committed 12 days ago

Verified b879589a

Fix BF16 owned-fragment gradient reconstruction on current master

yh0903 committed 13 days ago

bd52453e

Keep Muon's update at full size under fp16 loss scaling (#8655)

alanhuangyoo committed 13 days ago

Verified 3509874e

Merge upstream master for BF16 gradient lifecycle review

yh0903 committed 13 days ago

f1a000c4

Nightly full-suite run with classified failure triage (#8606)

delock committed 13 days ago

Verified 865fde4f

AutoEP: one program per token in the fused weighted restore forward (#8678)

yh0903 committed 13 days ago

Verified 11b24d21

Restore a shard from the map the layer published (#8622)

Achyuthan-S committed 14 days ago

Verified a80df94b

Use expert gradients, not expert weights, in the FP16_Optimizer MoE grad norm (#8635)

yh0903 committed 14 days ago

Verified 1b60c549

Validate the AutoSP SP/DP mesh when registering process groups (#8674)

jinyouzhi committed 14 days ago

Verified 084e5f6e

Check AutoSP sequence length and head count divisibility (#8675)

jinyouzhi committed 14 days ago

Verified 245c3a98

Fix FPDT K/V gradient buffers under grouped-query attention (#8665)

jinyouzhi committed 14 days ago

Verified ece19cfe

Reject amsgrad in DeepSpeedCPUAdam instead of ignoring it (#8694)

vineethsaivs committed 14 days ago

Verified 53e336cd

Fix ZeRO-3 Muon all-gather for mixed parameter shapes (#8628)

jinyouzhi committed 14 days ago

Verified c9e1b4d2

Consolidate PRs for docs and typos (#8689)

hwchen2017 committed 15 days ago

Verified 62faee36

Skip fp16-config tests on accelerators without fp16 support (#8684)

delock committed 15 days ago

Verified df00e0ba

fix(rollout): forward top-k sampling configuration (#8676)

nathon-lee committed 15 days ago

Verified 729d5aeb

Older