microsoft/DeepSpeed

Pull Requests Commits

Add an opt-in fused weighted restore for AutoEP (#8326)

yh0903 committed 27 days ago

Verified 80f19f30

Stop the curriculum schedule starting below min_difficulty (#8334)

vineethsaivs committed 27 days ago

Verified 8e09ed2a

Keep the elasticity batch overrides out of the caller's config dict (#8329)

alanhuangyoo committed 27 days ago

Verified c7eed159

Replace cuda_graph assertion with explicit ValueError; fix custom_op type annotations (#8336)

chakshu-dhannawat committed 27 days ago

Verified 26b2d995

Broadcast elementwise flops from the trailing dimension (#8324)

vineethsaivs committed 27 days ago

Verified b8bfd260

Fix flops profiler counts for transposed convolutions (#8323)

vineethsaivs committed 27 days ago

Verified 7fabe473

Forward barrier device_ids to communication backends (#8312)

tohtana committed 27 days ago

Verified 177c5aae

Forward free_data through partition() instead of hardcoding True (#8305)

vineethsaivs committed 27 days ago

Verified 9666fe5f

Recognize Qwen3.5's RMSNorm variants in AutoTP module loading (#8306)

promptsmith1990 committed 27 days ago

Verified 8e64a09b

Make the WarmupCosineLR ratio flags reach the config (#8268)

vineethsaivs committed 27 days ago

Verified cdd206ff

fix: Only bind device id when needed, Fixes #8248 (#8269)

pengdurice committed 28 days ago

Verified 3bdabae0

DeepCompile: stabilize ZeRO-3 parameter guards (#8328)

tohtana committed 28 days ago

Verified 41890912

perf(rollout): profile prefill and decode forwards (#8350)

nathon-lee committed 28 days ago

Verified 1c5cc6bf

remove the dead non-Triton attention path from TritonSelfAttention (#8349)

Anai-Guo committed 28 days ago

Verified 5b6da2be

Add an opt-in DeepEP transport for the AutoEP expert all-to-all (#8213)

yh0903 committed 28 days ago

Verified 534dc0e7

Add macOS (MPS) CI workflow and a torch floor check for the MPS accelerator (#8335)

PKUWZP committed 29 days ago

Verified ba3246de

feat(cpu-adam): add ARM SVE update kernel (#8365)

xylian86 committed 29 days ago

Verified 020893ef

[AutoTP] Replace tp_shard process-wide globals with per-model AutoTPMeta (#8241)

delock committed 29 days ago

Verified 183c7f95

Fix Triton NFS detection crash when df wraps long device names (#8259)

atirna committed 30 days ago

Verified c389bae3

Add pin_empty helper for empty scratch destinations (#8332)

sfc-gh-truwase committed 30 days ago

Verified 725811f9

[tiled mlp] reshape instead of view (#8348)

stas00 committed 31 days ago

Verified 87d9ecd8

Fallback for unsupported Hybrid Engine policies (#8265)

LiRunGuo committed 31 days ago

Verified a36f78e7

Count each module object once when aggregating flops profiler totals (#8320)

ebarkhordar committed 31 days ago

Verified f4ee787d

Fix the seq-first Ulysses all2all output layout (#8317)

vineethsaivs committed 31 days ago

Verified cd5206af

Raise the per-tensor norms to norm_type when combining them (#8313)

vineethsaivs committed 31 days ago

Verified 37cf8f24

Add configurable dtype for ZeRO checkpoint export (#8318)

gaoxiaomo committed 33 days ago

Verified 92843ad7

Honor adam_w_mode in the CPU multi_tensor_adam binding (#8307)

PKUWZP committed 33 days ago

Verified c7cc64a9

Route ZeRO/SuperOffload pin sites through accelerator pin_memory (#8256)

sfc-gh-truwase committed 33 days ago

Verified eebe24a6

[AutoTP] Fix training lm_head routing (#8302)

gaoxiaomo committed 33 days ago

Verified 1b956e62

Fix ZeRO-3 synchronization during OPSD rollout (#8264)

LiRunGuo committed 33 days ago

Verified e40c681b

Newer Older