microsoft / DeepSpeed Public

2 Pull requests merged by 2 people

• Refactor MoE and Groups API to simplify model creation and mangement

  #1798 merged Feb 28, 2022

• Bump nokogiri from 1.12.5 to 1.13.3 in /docs

  #1794 merged Feb 28, 2022

2 Pull requests opened by 2 people

• 01 adam optimizer

  #1790 opened Feb 24, 2022

• Website posts improvements

  #1799 opened Feb 28, 2022

6 Issues opened by 6 people

• [BUG] DeepSpeed Inference with GPT-J using batches with padding gives wrong outputs

  #1797 opened Feb 27, 2022

• [REQUEST] Can I manually change global batch size with checkpoints?

  #1796 opened Feb 27, 2022

• [BUG] IndexError / Runtime Error with `torch.nn.TransformerEncoder`

  #1795 opened Feb 26, 2022

• [BUG] zero_to_fp32.py cannot convert the model

  #1793 opened Feb 25, 2022

• [BUG] save_checkpoint() missing some params

  #1789 opened Feb 24, 2022

• [BUG] CUDA error with INT 8 inference

  #1788 opened Feb 23, 2022

11 Unresolved conversations

Sometimes conversations happen on old items that aren’t yet closed. Here is a list of all the Issues and Pull Requests with unresolved conversations.

• Cannot reproduce performance from the paper for 50B and 100B models

  #1776 commented on Feb 23, 2022 • 3 new comments

• Performance Degradation with ZERO Stage 3

  #1069 commented on Feb 25, 2022 • 3 new comments

• why cpu_checkpointing can't work?

  #522 commented on Feb 28, 2022 • 2 new comments

• [BUG] Support for MoE model inference

  #1743 commented on Feb 23, 2022 • 1 new comment

• [BUG] `get_model_profile()` for `nn.Upsample(scale_factor=2)` does not work

  #1701 commented on Feb 24, 2022 • 1 new comment

• [Potential contribution] Minibatch trimming (curriculum learning method)

  #1539 commented on Feb 24, 2022 • 1 new comment

• Unable to find hostfile

  #1783 commented on Feb 25, 2022 • 1 new comment

• Suggestion: Support CPU-only environments/remove CUDA requirement

  #1279 commented on Feb 27, 2022 • 1 new comment

• AMD support

  #1430 commented on Feb 28, 2022 • 0 new comments

• DataLoader Length Fix

  #1718 commented on Feb 23, 2022 • 0 new comments

• Optimizer state for only trainable parameters

  #1780 commented on Feb 25, 2022 • 0 new comments