microsoft/DeepSpeed

Pull Requests Commits

Add fused triton kernel for swiglu (#8244)

hwchen2017 committed 12 days ago

Verified 24a41afa

Route FPDT and checkpoint writer pins through accelerator pin_memory (#8257)

sfc-gh-truwase committed 12 days ago

Verified 310eb8cb

[AutoTP] Complete uneven sharding and universal checkpoint support (#8185)

jinyouzhi committed 13 days ago

Verified aa3914df

Keep parameter dtype through ZeRO-3 weight quantization (#8215)

adityasingh2400 committed 13 days ago

Verified 6098f799

Fix docstring Args entries that name a parameter the function does not take (#8223)

darkdi committed 13 days ago

Verified 32d51a18

Add CUDA graph support for HybridEngine generation (#8271)

PKUWZP committed 14 days ago

Verified 80c8e5ba

Fix ZeRO-3 crash in AutoTP universal-checkpoint metadata (#8270)

PKUWZP committed 14 days ago

Verified 313ce47b

NCCL and other backend PG timeout 30min -> 10min (#8253)

stas00 committed 15 days ago

Verified 24402c3b

Fix int32 overflow in Triton grouped-GEMM expert offset (#8261)

PKUWZP committed 16 days ago

Verified c5331bc1

Add compile.offload_activation_pin_memory for DeepCompile activation offload (#8258)

sfc-gh-truwase committed 17 days ago

Verified c952d92a

Add native (DeepNVMe) host-memory pinning backend for accelerators (#8211)

sfc-gh-truwase committed 18 days ago

Verified aa0e91b9

Add configurable sum gradient reduction (#8232)

zupengwang committed 18 days ago

Verified 1cae4920

Fix DeepCompile last use for unconsumed waits (#8254)

tohtana committed 18 days ago

Verified 9bd89f9d

(1/2) Implementing Compiler Pass for AutoTP (#8204)

therealnaveenkamal committed 19 days ago

Verified 79046032

Upgrade pip in Python install-smoke jobs (#8247)

tohtana committed 20 days ago

Verified 22969fa7

[DeepCompile] Fix staticmethod handling on Python 3.9 (#8240)

baremetaldevx86 committed 21 days ago

Verified 1d580d65

[DeepCompile] Fix KeyError on frozen parameters in ZeRO-3 (#8214)

roycho96 committed 21 days ago

Verified 04b4b0cb

Keep the elastic batch size within max_train_batch_size (#8237)

vineethsaivs committed 21 days ago

Verified 2cfebbda

Implement the documented per-param-group lists in OneCycle (#8201)

vineethsaivs committed 21 days ago

Verified 8cae9d28

Update version.txt after 0.19.5 release (#8243)

loadams committed 21 days ago

Verified 66f9a364

Return a copy from OnebitLamb.get_lamb_coeffs (#8227)

vineethsaivs committed 22 days ago

Verified 5dceeeab

Fix checkpoint rank selection for Ulysses sequence parallelism (#8226)

sfc-gh-thonguyen committed 22 days ago

Verified b39e07a7

Unmanaged gradient accumulation: ZeRO offload support (#8225)

sfc-gh-truwase committed 23 days ago

Verified cf443004

Split native host pinning into standalone pin_memory op (#8236)

sfc-gh-truwase committed 24 days ago

Verified 5ad9a978

Fix compilation error for torch 2.12+ (#8238)

hwchen2017 committed 24 days ago

Verified a2f8d9ea

Fix AutoEP ZeRO-1/2 universal conversion (#8198)

tohtana committed 25 days ago

Verified da066407

fix(zero3): async grad offload + pinned offload buffers by default (#8207)

delock committed 26 days ago

Verified 7dfa67ae

Fix ZeRO++ secondary shard copy for small params (#8210)

ZyEng-Art committed 26 days ago

Verified 2d4488ef

Do not let one op builder's compatibility probe break importing deepspeed (#8216)

adityasingh2400 committed 26 days ago

Verified 646e491c

Update actions versions for workflows. (#8228)

loadams committed 26 days ago

Verified c6e72794

Newer Older