microsoft/DeepSpeed

Pull Requests Commits

ci: retrigger sandbox probe

delock committed 1 day ago

9275430c

ci: probe host GPU driver version in sandbox test logs

delock committed 1 day ago

52d7f4d7

[bug fix] ZeRO2: preserve fp32 grad accumulation precision in partition storage (#8749)

stas00 committed 1 day ago

Verified 3b809329

Use C++20 for XPU op builders on torch 2.12+ (#8803)

jinyouzhi committed 2 days ago

Verified dcd8595c

AutoEP: do not slice DeepEP buffers that already hold exactly the arrived rows (#8747)

yh0903 committed 2 days ago

Verified b2047562

Add a pr-checklist agent skill and an AI contribution policy (#8744)

PKUWZP committed 3 days ago

Verified 1c862f7d

Add an opt-in fused gate+up grouped GEMM for AutoEP experts (#8733)

yh0903 committed 3 days ago

Verified 3a656d3e

Replay AutoEP's DeepEP dispatch layout in non-reentrant checkpoint recompute (#8731)

yh0903 committed 3 days ago

Verified 9ec13539

Add an opt-in chunked causal-LM loss that keeps no full-vocabulary FP32 tensor (#8653)

yh0903 committed 3 days ago

Verified cd4cb971

[AutoSP] Rebuild query-sharded attention masks before SDPA (#8761)

jinyouzhi committed 3 days ago

Verified ee222bf3

[AutoSP] Fix tensor truthiness in get_node_shape_meta (#8760)

jinyouzhi committed 3 days ago

Verified 52fd6781

[AutoSP] Honor seq_dim when AutoSP shards inputs (#8720)

jinyouzhi committed 3 days ago

Verified a13f6b04

[checkpoint] Refuse a map whose pieces write the same shard address (#8624)

0z5a committed 4 days ago

Verified d371f33c

[Bug Fix] Fix CPUAdagrad updates with nonzero weight decay (#8781)

hwchen2017 committed 4 days ago

Verified e759fe43

Preserve non-floating buffers in ZeRO checkpoint conversion (#8515)

aprylewu committed 5 days ago

Verified a20acd1d

[UC] Plan a scale-1 affine shard-to-shard transfer without rebuilding the parameter (#8623)

0z5a committed 5 days ago

Verified ae1b183f

Log the skipped optimizer step on gradient overflow (#8417)

sfc-gh-abkulkarni committed 5 days ago

Verified 500997a4

Add an opt-in fused Triton RMSNorm for Llama-style Hugging Face models (#8645)

yh0903 committed 5 days ago

Verified 3ca100d6

AutoEP + AutoTP folding under ZeRO-3 (#8670)

pengdurice committed 5 days ago

Verified 7d77213d

Avoid circuler import caused by Hybrid Engine and OPT model (#8737)

tohtana committed 5 days ago

Verified f47a963d

Reduce unnecessary CI work for documentation changes and consolidate agent instructions in AGENTS.md (#8794)

hwchen2017 committed 5 days ago

Verified bf8f489a

Guard the shm inference all-reduce on the host ISA (#8701)

delock committed 5 days ago

Verified 18f92828

Use default fp32 tolerances in AutoTP training tests (#8796)

hwchen2017 committed 5 days ago

Verified 967803b0

feat(rollout): add OPSD response token loss primitives (#8677)

nathon-lee committed 5 days ago

Verified f1b516b6

Fix DataAnalyzer.run_map crashing when specific_threads is set (#8758)

Rakshit-gen committed 6 days ago

Verified b7f51df9

Keep the buffers a model names for fp32 in fp32 under ZeRO-3 (#8699)

pengdurice committed 6 days ago

Verified 7b911151

[AutoTP] Support tied embeddings for row-parallel output-head training (#8680)

jinyouzhi committed 6 days ago

Verified 3c6a5ca9

Assert offload-state removal only when offload leaves the accelerator (#8738)

delock committed 6 days ago

Verified 1707800a

Stream AWS fallback logs while commands run (#8783)

tohtana committed 6 days ago

Verified ff85fbfc

Release each BF16 gradient once immediate_grad_update has accumulated it (#8654)

yh0903 committed 6 days ago

Verified bc1ad320

Older