Pull Requests microsoft/DeepSpeed

Recognize a norm by its shape, not by whether its class name was listed

#8479 opened 2026-09-10 19:59 by alanhuangyoo

Count each expert once in the fp32 gradient-clipping norm

#8478 opened 2026-09-10 19:07 by alanhuangyoo

Carry the affine scale on the replicated map, not the split

#8477 opened 2026-09-10 17:40 by Achyuthan-S

Fix AutoEP global L2 clipping for sharded expert ownership

#8476 opened 2026-09-10 16:43 by gss10282023

Reject a non-positive out-token limit at config time

#8473 opened 2026-09-10 14:54 by alanhuangyoo

Make add_tuning_arguments' 1Cycle flags reach OneCycle

#8472 opened 2026-09-10 10:28 by alanhuangyoo

fix: track FusedAdam bias correction per parameter

#8471 opened 2026-09-10 09:36 by MrCapricornLiu

Support dynamic parameter offload for ZeRO-3 inference

#8468 opened 2026-09-10 00:35 by zupengwang

A zero-element parameter loses its shape when it is bound to a flat buffer

#8467 opened 2026-09-09 14:14 by alanhuangyoo

fix(inference): break circular import in ops.transformer.inference

#8465 opened 2026-09-09 02:50 by chakshu-dhannawat

Fix Muon optimizer under ZeRO CPU offload and bound gather buffers

#8464 opened 2026-09-08 16:33 by jinyouzhi

Checkpoint the RNG the curriculum sampler actually draws from

#8460 opened 2026-09-08 12:06 by alanhuangyoo

[AutoSP] Fix compiler correctness

#8457 opened 2026-09-08 10:21 by jinyouzhi

Say when ZenFlow will never re-select its important columns

#8456 opened 2026-09-08 06:45 by alanhuangyoo

Rebuild the full argument list when partitioning activations

#8455 opened 2026-09-08 06:44 by vineethsaivs

fix(inference): reject non-positive max_out_tokens at config validation

#8454 opened 2026-09-08 05:17 by chakshu-dhannawat

[Phase 2] Add NEON SIMD path for CPU Adam on AArch64

#8453 opened 2026-09-08 04:47 by PKUWZP

Sort checkpoint shard files by numeric rank

#8452 opened 2026-09-08 02:44 by ebarkhordar

Add opt-in AutoEP Python GC policy

#8451 opened 2026-09-07 19:56 by yh0903

docs(runtime): document max_norm and use_graph in clip_tensors_by_global_norm

#8450 opened 2026-09-07 16:41 by simpleqt

docs: point CONTRIBUTING's installation link at the README section

#8449 opened 2026-09-07 16:20 by simpleqt

docs(sparse_attention): fix copy-pasted BigBird return line in sliding-window layout

#8448 opened 2026-09-07 16:20 by simpleqt

fix(checkpointing): remove identical if/else arms and leftover debug print in WriterFactory

#8446 opened 2026-09-06 17:35 by simpleqt

fix(zenflow): restore assert and remove duplicated identical branch in gradient copy

#8445 opened 2026-09-06 17:34 by simpleqt

DeepSpeed cannot start without mpi4py on a machine with no launcher

#8444 opened 2026-09-06 17:26 by alanhuangyoo

Muon silently discards the param groups it is given

#8440 opened 2026-09-06 13:51 by alanhuangyoo

[muon] Per-head Muon for linear-attention layers: read head geometry from the owning module

#8436 opened 2026-09-06 11:30 by alanhuangyoo

[muon] Keep the momentum out of steps the loss scaler discards

#8435 opened 2026-09-06 10:20 by alanhuangyoo

Fix SequenceTiledCompute backward for empty trailing shards

#8434 opened 2026-09-06 10:02 by taking-lying-flat

[muon] Reconcile the momentum dtype when a checkpoint is restored

#8433 opened 2026-09-06 10:01 by alanhuangyoo