deepspeedai / deepspeedai/DeepSpeed
Question about Ulysses and loss agregation
Open
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 43.1k
- Forks
- 5k
- Avg merge
- 4d 15h
- Merged PRs (30d)
- 112
Description
HI,
I am using Ulysses Attention and DeepSpeed Zero3 optimizer for DPO training.
My question what is the right way of loss aggregation?
When one trains model with CrossEntropyLoss, each rank yields loss for each own sequence subset.
But when on trains model with DPO loss, we have only loss for the whole example.
What is right way to deal with it?
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
No files, tests, or entry points are mentioned. Start by reviewing the Ulysses and DeepSpeed ZeRO-3 training integration alongside the DPO loss path, then clarify the expected loss aggregation behavior before defining any documentation or code changes.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python, pytorch
- Domain
- distributed-systems, machine-learning
- Issue type
- Documentation
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 15/100