deepspeedai / deepspeedai/DeepSpeed

[BUG] MOE: Loading experts parameters error when using expert parallel.

Open
#6,589 1 comment 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
I have a model with 60 experts, and I am training the experts in parallel on two GPUs. Theoretically, GPU0 should load the parameters of the first 30 experts, while GPU1 should load the parameters of the last 30 experts. However, I found that both GPUs are loading the parameters of the first 30 experts . How should I modify this?

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 reproducing the reported setup with 60 experts and expert parallelism across two GPUs, then trace the expert-parameter loading path to determine why both ranks select the first 30 experts. Done means GPU0 loads experts 0–29 and GPU1 loads experts 30–59, with a regression test covering the two-GPU case.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
distributed-systems, machine-learning
Issue type
Bug
Difficulty
4/5
Estimated time
3-5 days
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
25/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.