Commits deepspeedai/DeepSpeed

Reduce blockDim in fake_quantize_kernel for improved SM occupancy (#8115)

flutist committed 24 days ago

Verified 000ca191

fix(autotp): when using autotp, ignore the consistency of certain data within the tp_group. (#8125)

1787648106 committed 24 days ago

Verified 53a2ac44

Add EXAONE 4.5 model support for Inference V2 (#8121)

Bias92 committed 25 days ago

Verified d36079f9

Guard WarmupCosineLR against total_num_steps == warmup_num_steps (ZeroDivisionError) (#8142)

vineethsaivs committed 25 days ago

Verified 2163a013

Fix CommsLogger.stop_profiling_comms not disabling global profiling (#8137)

vineethsaivs committed 25 days ago

Verified 10db6919

ZeRO 1/2: wait on all IPG-bucket producer streams in average_tensor (#8061) (#8080)

arunshar committed 26 days ago

Verified 711bd11e

Make DCO workflow Probot compatible (#8110)

tohtana committed 27 days ago

Verified 46c9e20f

Add AutoEP + AutoTP parallel folding (#8064)

tohtana committed 27 days ago

Verified 104193b4

Add Hybrid Engine rollout in DeepSpeed to support On-Policy Distillation (OPSD) Trainer (#8027)

PKUWZP committed 28 days ago

Verified 429e2ad2

feat(types): add type hints to public comm API functions (#8118)

NovusEdge committed 28 days ago

Verified 9b106c49

doc: Remove suggestion to build extensions in parallel (#7899)

Flamefire committed 28 days ago

Verified 05a481b8

Fix ZeRO-3 autocast gather with mixed parameter dtypes (#8113)

tohtana committed 28 days ago

Verified 195255e7

Fix ZeRO-3: use per-param dtype for quantize-scale buffer in _allgather_params_coalesced (#8128)

roycho96 committed 29 days ago

Verified b079c309

Add input validation for save_checkpoint and warmup LR schedulers (#8126)

jahnavi-yelamanchi committed 30 days ago

Verified 54ebdec6

Fix Python 3.9 import-time TypeError in AutoEP ep_router (#8119)

vineethsaivs committed 33 days ago

Verified ae1dba54

Fix FlopsProfiler crash when dp_world_size is None under sequence parallelism (#8122)

vineethsaivs committed 33 days ago

Verified 6e38d56f

fix: add buffer-length check in shm.cpp (#8082)

orbisai0security committed 33 days ago

Verified 7720656e

feat(zenflow): run the overlapped CPU optimizer in a native process (#8058)

Antlera committed 33 days ago

Verified 1ddbe484

Fix minor comment/docstring typos in runtime and inference modules (#8046)

nathon-lee committed 38 days ago

Verified 0b987c0e

Add optional torchembed RoPE backend to apply_rotary_pos_emb (#8052)

py-ai-dev committed 38 days ago

Verified 5ca819bc

Fix DeepCompile profiling memory cleanup (#8106)

tohtana committed 38 days ago

Verified 131779cf

fix: log eigenvalue monitor values (#8049)

he-yufeng committed 38 days ago

Verified 32d009bd

Feat: zero3 deprecate elastic checkpoint (#8099)

nathon-lee committed 38 days ago

Verified ac5c49f5

Fix eigenvalue parsing for compression-only quantize configs (#8057)

sowndappan5 committed 38 days ago

Verified ea3d65ea

[DeepCompile] fix gather params in dynamo skipped frames for ZeRO3 (#8059)

XAheli committed 39 days ago

Verified 1e85ce2b

Avoid CUDA context initialization during op compatibility checks at import (#8078)

Achyuthan-S committed 41 days ago

Verified 5aef6d88

Don't swallow KeyboardInterrupt/SystemExit in NPUOpBuilder (#8096)

ajinkyajawale14499 committed 41 days ago

Verified 10f76c2e

Add numerical-correctness test for Muon under ZeRO-1/2 (#8091)

whycoming committed 41 days ago

Verified 0ecb987d

Fix DeepCompile profile metadata backfill (#8094)

tohtana committed 42 days ago

Verified a33655cb

Reject Muon optimizer with reduce_scatter in ZeRO-1/2 (#8090)

whycoming committed 42 days ago

Verified d714251c