-
Notifications
You must be signed in to change notification settings - Fork 5k
Pull requests: deepspeedai/DeepSpeed
Author
Label
Projects
Milestones
Reviews
Assignee
Sort
Pull requests list
implement register_host_memory and unregister_host_memory for supa
#8537
opened Sep 16, 2026 by
frozenleaves
Contributor
Loading…
Cache the Modal sandbox install chain in image layers
#8536
opened Sep 16, 2026 by
delock
Collaborator
Loading…
Deprecate leftovers of compression library
#8535
opened Sep 16, 2026 by
hwchen2017
Contributor
Loading…
Normalize Gram Newton-Schulz inputs before casting to fp16
#8533
opened Sep 16, 2026 by
vineethsaivs
Contributor
Loading…
Fix ZeRO-3 checkpoint conversion with --debug
#8530
opened Sep 15, 2026 by
vineethsaivs
Contributor
Loading…
Keep Random-LTD sequence lengths above the configured minimum
#8529
opened Sep 15, 2026 by
vineethsaivs
Contributor
Loading…
Keep tensor scheduler bounds fixed across LR updates
#8528
opened Sep 15, 2026 by
vineethsaivs
Contributor
Loading…
Support Muon in BF16_Optimizer with FP32 gradient accumulation
#8526
opened Sep 15, 2026 by
0z5a
Contributor
Loading…
fix(data_pipeline): default device to active accelerator in DistributedDataAnalyzer
#8525
opened Sep 15, 2026 by
li-lizhe
Loading…
fix(zenflow): use optimizer_z3.device instead of hardcoded cuda
#8524
opened Sep 15, 2026 by
li-lizhe
Loading…
Fix HybridEngine OPT generation with modern cache interfaces
#8523
opened Sep 15, 2026 by
0z5a
Contributor
Loading…
instrument_w_nvtx drops the wrapped function's name, docstring and signature
#8521
opened Sep 15, 2026 by
ebarkhordar
Contributor
Loading…
Fill curriculum batches across repeated cluster shuffles
#8520
opened Sep 15, 2026 by
vineethsaivs
Contributor
Loading…
Accept gradient iterators in global norm helpers
#8518
opened Sep 14, 2026 by
vineethsaivs
Contributor
Loading…
Restore memory-mapped datasets when unpickling
#8517
opened Sep 14, 2026 by
vineethsaivs
Contributor
Loading…
Skip absent buffers when loading AutoTP checkpoints
#8516
opened Sep 14, 2026 by
vineethsaivs
Contributor
Loading…
Preserve non-floating buffers in ZeRO checkpoint conversion
#8515
opened Sep 14, 2026 by
aprylewu
Loading…
ZeRO-1/2: add bounded quantized weight all-gather
#8511
opened Sep 14, 2026 by
adenzhou1350
Loading…
Count the operator spelling of add and mul in the flops profiler
#8509
opened Sep 14, 2026 by
ebarkhordar
Contributor
Loading…
Preserve ZeRO-3 release output metadata in DeepCompile
#8508
opened Sep 14, 2026 by
tohtana
Collaborator
Loading…
Count PackedSequence cell steps in the flops profiler RNN hook
#8507
opened Sep 14, 2026 by
ebarkhordar
Contributor
Loading…
Previous Next
ProTip!
Updated in the last three days: updated:>2026-09-13.