microsoft/DeepSpeed

Pull Requests Commits

hwchen2017 committed 1 hour ago

cf61b715

Deprecate sparse attention

hwchen2017 committed 1 hour ago

ffeb464b

Deprecate unused DeepSpeed features (#8490)

sfc-gh-truwase committed 7 hours ago

Verified 71d316d6

[muon] Reconcile the momentum dtype when a checkpoint is restored (#8433)

alanhuangyoo committed 21 hours ago

Verified b726f4ed

Carry the affine scale on the replicated map, not the split (#8477)

Achyuthan-S committed 23 hours ago

Verified bce0adf0

Fix ZeRO parameter alignment for grouped_mm (#8277)

fwerkor committed 1 day ago

Verified 080eb6b0

feat(rollout): add continuous batching generation prototype (#8368)

nathon-lee committed 1 day ago

Verified e5887a35

Fix universal checkpoint resume across AutoTP sizes (#8474)

0z5a committed 1 day ago

Verified e76a2805

Fix comms logger KeyError when log_name is omitted (#8267)

jinyouzhi committed 1 day ago

Verified bd6e9681

Describe universal checkpoint shards as affine maps (#8385)

Achyuthan-S committed 1 day ago

Verified 11909466

Split DeepCompile ZeRO-3 memory scheduler (#8233)

tohtana committed 1 day ago

Verified 38561463

[MPS] Update C++ Standard in CPUAdamBuilder (#8466)

jinyouzhi committed 1 day ago

Verified bb7ad0d8

Avoid collective token preparation for AutoEP DeepEP (#8423)

yh0903 committed 2 days ago

Verified 29d0abbc

Muon is silently disabled under ZeRO-3 when the model is built with zero.Init (#8438)

alanhuangyoo committed 2 days ago

Verified 0166cc87

[AutoEP]Fix optimizer and replaced MOE parameter mismatch (#8377)

pengdurice committed 2 days ago

Verified 5daeffeb

Filter --include against the real slots, not against itself (#8239)

vineethsaivs committed 2 days ago

Verified 3b1c14ac

Add an opt-in ZeRO-1/2 gradient norm fast path (#8331)

yh0903 committed 2 days ago

Verified f992dc24

Muon runs no Newton-Schulz at ZeRO stage 0, the default: run it (#8442)

alanhuangyoo committed 3 days ago

Verified cbd303eb

Stop the debug name maps from pinning the model they snapshot (#8356)

alanhuangyoo committed 4 days ago

Verified e9680c7d

Read rope_theta from rope_parameters in the Llama injection policy (#8341)

alanhuangyoo committed 4 days ago

Verified 6474bc55

[Workflow] run modal GPU workflows only from the merge queue (#8412)

delock committed 4 days ago

Verified 1e90aa9f

Split the modal CI budget into acquisition and test phases (#8404)

delock committed 5 days ago

Verified 666720b4

fix(lr_schedules): make --lr_range_test_staircase an opt-in flag (#8337)

Anai-Guo committed 5 days ago

Verified 8d4074cb

Fix AutoTP + deep compile collectives silently drop when AC is on (#8355)

pengdurice committed 5 days ago

Verified 92269f43

Reshape instead of view in TiledFusedLogitsLoss (#8362)

alanhuangyoo committed 5 days ago

Verified 71199367

op_builder: use C++20 for nvcc on CUDA 13+ (#8422)

bheu committed 5 days ago

Verified abf3dfa4

Preserve AutoEP score correction bias buffers (#8369)

poorpaper committed 6 days ago

Verified 32110a95

[Workflow] Raise modal CI timeouts to absorb slower sandbox provisioning (#8403)

delock committed 8 days ago

Verified 05daf059

[Ulysses] Carry the KV head count per DistributedAttention (#8316)

alanhuangyoo committed 8 days ago

Verified d4ed1f14

Use device names, not rank ids, for device placement in test helpers (#8397)

delock committed 8 days ago

Verified 4fd3c825

Older