deepspeed
https://github.com/microsoft/deepspeed
Python
DeepSpeed is a deep learning optimization library that makes distributed training easy, efficient, and effective.
Triage Issues!
When you volunteer to triage issues, you'll receive an email each day with a link to an open issue that needs help in this project. You'll also receive instructions on how to triage issues.
Triage Docs!
Receive a documented method or class from your favorite GitHub repos in your inbox every day. If you're really pro, receive undocumented methods or classes and supercharge your commit history.
Python not yet supported21 Subscribers
Add a CodeTriage badge to deepspeed
Help out
- Issues
- Fix DynamicLossScaler never lowering the scale when fp16 hysteresis is 0 or less
- ZeRO-3: fix stale and lost gradients of unused parameters
- Merge final_linear shards when loading a Megatron checkpoint on fewer TP ranks
- Sum the OPSD JSD loss in float32 for bf16 and fp16 logits
- [BUG] zero.Init casts tensors created with an explicit dtype=torch.float (RoPE inv_freq computed in bf16)
- [BUG] OneCycle applies the minimum momentum to the first Adam update
- [BUG] Re-exporting a ZeRO checkpoint can leave a stale index that loads old weights
- Initialize OneCycle momentum at the upper bound
- Retire stale checkpoint artifacts after replacing a ZeRO export
- ZeRO-1/2 gradient accumulation: remaining test coverage
- Docs
- Python not yet supported