-
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
Restore full training state in checkpoint writer engines
#8502
opened Sep 13, 2026 by
ai-yang
Loading…
Fix ZeRO-3 checkpoint tag selection during weight reconstruction
#8501
opened Sep 13, 2026 by
ai-yang
Loading…
docs: add Simplified Chinese translation (README.zh.md)
#8498
opened Sep 13, 2026 by
JasonYeYuhe
Loading…
3 tasks done
feat(rollout): align decode fronts and trim dead zones
#8497
opened Sep 13, 2026 by
nathon-lee
Contributor
Loading…
Keep ZenFlow Adam step counters in float32
#8496
opened Sep 13, 2026 by
vineethsaivs
Contributor
Loading…
Preserve each Adam group's beta2 in OneCycle
#8495
opened Sep 13, 2026 by
vineethsaivs
Contributor
Loading…
feat(rollout): add continuous batching profiling
#8494
opened Sep 12, 2026 by
nathon-lee
Contributor
Loading…
Remove triton compatibility check for fp_quantizer
#8492
opened Sep 12, 2026 by
hwchen2017
Contributor
•
Queued
[bugfix] _DimZeroAllToAll silently fails to send grads under torch.compile
#8491
opened Sep 11, 2026 by
stas00
Collaborator
Loading…
Refresh the whole captured input before a CUDA graph replay
#8488
opened Sep 11, 2026 by
vineethsaivs
Contributor
Loading…
Make wait() idempotent on AllGatherHandle and NoGatherHandle
#8487
opened Sep 11, 2026 by
pengdurice
Contributor
•
Draft
Give Muon's momentum the dtype of the gradient it is combined with
#8483
opened Sep 11, 2026 by
alanhuangyoo
Contributor
Loading…
Fixed memory leak in loss backward with AC & zero3 & single-model-multi-branch network
#8482
opened Sep 11, 2026 by
supermeng
Loading…
Recognize a norm by its shape, not by whether its class name was listed
#8479
opened Sep 10, 2026 by
alanhuangyoo
Contributor
Loading…
Count each expert once in the fp32 gradient-clipping norm
#8478
opened Sep 10, 2026 by
alanhuangyoo
Contributor
Loading…
Fix AutoEP global L2 clipping for sharded expert ownership
#8476
opened Sep 10, 2026 by
gss10282023
Loading…
Reject a non-positive out-token limit at config time
#8473
opened Sep 10, 2026 by
alanhuangyoo
Contributor
Loading…
Make add_tuning_arguments' 1Cycle flags reach OneCycle
#8472
opened Sep 10, 2026 by
alanhuangyoo
Contributor
Loading…
fix: track FusedAdam bias correction per parameter
#8471
opened Sep 10, 2026 by
MrCapricornLiu
Loading…
Support dynamic parameter offload for ZeRO-3 inference
#8468
opened Sep 10, 2026 by
zupengwang
Contributor
Loading…
Previous Next
ProTip!
Add no:assignee to see everything that’s not assigned.