deepspeedai / deepspeedai/DeepSpeed

Data Loading for DeepSpeed Ulysses and Data Parallelism

Open
#5,582 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

bug training
Dominant language
Python
Stars
43.1k
Forks
5k
Avg merge
4d 15h
Merged PRs (30d)
112

Description

Describe the bug
After reading the source code here, here and some existing issues such as https://github.com/microsoft/DeepSpeed/issues/4359#issuecomment-1726166408, it seems like if we want to use Ulysses with Data Parallelism, we need to modify the data loader such that for each sequence, each GPU needs to load 1/P (assume we have P GPUs) of the sequence.

If my understanding is incorrect, it would be great if you could let me know what Ulysses actually does and, if needed, how we should modify the data loader to support it. The blog post and paper are not very clear about the implementation details.

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start by reading deepspeed/sequence/layer.py at the linked lines and trace how Ulysses interacts with data parallelism. Check the existing issue discussion referenced in the report; done means establishing whether the data loader needs changes and documenting the required behavior or implementation scope.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
distributed-systems, machine-learning
Issue type
Bug
Difficulty
5/5
Estimated time
Over a week
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
20/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.