deepspeedai/DeepSpeed

Pull Requests Commits

Fix unit tests failure on CPU that support fp16

hwchen2017 committed 17 minutes ago

87767379

[Ulysses] Add opt-in uneven-shard rejection and FPDT padding (#9059)

jschneider-canva committed 18 hours ago

Verified 663de973

Reuse continuous rollout graphs and release captured resources on shutdown (#9009)

jinyouzhi committed 1 day ago

Verified f94aa5c5

Fused CPU Adam step (#8841)

pengdurice committed 1 day ago

Verified 84481c33

Update version.txt post-0.20.0 release (#9073)

loadams committed 1 day ago

Verified a75ab891

Update version.txt before next release (bump minor version) (#9072)

loadams committed 1 day ago

Verified 3202612e

Keep the parameters a model names for fp32 in fp32 under ZeRO-3 (#8843)

pengdurice committed 1 day ago

Verified 3fe2d950

Add deepspeed.models.glm_moe_dsa: TileLang sparse attention for GLM-5.2 (#8659)

pengdurice committed 1 day ago

Verified 18e005aa

Place restored step for fused and capturable optimizers on universal checkpoint load (#9056)

dundysm committed 2 days ago

Verified 988b2e31

Use the accelerator device in TestZero3SynchronizedEarlyStop (#9058)

dundysm committed 2 days ago

Verified dfa7d348

[AffineIR] Move shared affine index and box geometry into affine_geometry (#9010)

jinyouzhi committed 2 days ago

Verified 84679e9d

[checkpoint] Batch replica selection for affine transfer plans (#8936)

0z5a committed 2 days ago

Verified 2d02194a

klxpu: enable cuda graph capture (#9042)

GuCarpenter committed 2 days ago

Verified de6fcf18

Move test_klxpu_accelerator.py to tests/unit/v1/accelerator/ (#8867)

GuCarpenter committed 2 days ago

Verified c289ad62

Treat an empty subset selection as success in modal-torch-latest (#9053)

GuCarpenter committed 3 days ago

Verified b9100c54

Size HybridEngine rollout KV caches from local prefill shapes under AutoTP (#9008)

jinyouzhi committed 3 days ago

Verified 2dabbbe3

Synchronize ZeRO-3 rollout early stopping (#8390)

LiRunGuo committed 3 days ago

Verified 8a6b53fe

ZeRO-2 CPU offload: accumulate in the compact partition between requests (#9044)

sfc-gh-truwase committed 3 days ago

Verified f9fe4df3

Update an affiliation in COMMITTERS.md (#9050)

tohtana committed 3 days ago

Verified 51d86659

[NPU] Fix the fused LAMB step and inference int4/int8 dequantization on Ascend (#9043)

VenusTZZ committed 4 days ago

Verified f7a71777

Use C++20 for CUDA/ROCm op builders on torch 2.12+ (#9041)

VenusTZZ committed 4 days ago

Verified cbed412d

Code cleaning: remove commented out debug prints and code (#9036)

hwchen2017 committed 5 days ago

Verified 6f96b4a0

Fix memory leak in engine and zero optimizer (#9034)

hwchen2017 committed 5 days ago

Verified f32e164b

Persist ZeRO-3 HybridMegatron QKV layout changes (#8392)

LiRunGuo committed 5 days ago

Verified 988c88a0

ZeRO-1/2: one CPU-offload accumulator per group; zero idle gradients (#9018)

sfc-gh-truwase committed 5 days ago

Verified a4c989b7

[VocabTP] Add vocabulary-parallel generalized JSD with Triton kernels (#8958)

jinyouzhi committed 6 days ago

Verified da79d744

feat(rollout): non-greedy sampling on hybrid engine graph decode path (#8939)

anyadontfly committed 6 days ago

Verified 02f79784

Treat an all-deselected selection as success in cpu-torch-latest-v1-multirank (#9032)

delock committed 6 days ago

Verified 57ec847d

Fix routed-expert gradient clipping for AutoEP + AutoTP folding (#8823)

0z5a committed 7 days ago

Verified a7ebdf96

ZeRO-3 NVMe offload: fix "param already assigned swap buffer" in the parameter prefetch (#9014)

xylian86 committed 7 days ago

Verified 6c7a6e0e

Older