deepspeedai / deepspeedai/DeepSpeed
[QST] MoE auxiliary loss
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 43.1k
- Forks
- 5k
- Avg merge
- 4d 15h
- Merged PRs (30d)
- 112
Description
Hello!
My understanding is that the gate layer implements Algorithm 1 of GShard; however, our auxiliary loss computation seems to deviate from the algorithm; please help me understand.
That is, we compute l_aux here by sum(me * ce) * num_experts, while line 13 of the algorithm specifies mean(me * ce) or, equivalently, sum(me * ce) / num_experts.
Why do we deviate there?
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
Start in deepspeed/moe/sharded_moe.py at line 230 and compare the l_aux calculation with line 13 of Algorithm 1 in the linked GShard paper. Trace the gate-layer inputs to determine whether the scaling difference is intentional. Done means documenting the rationale or identifying the required correction, with any relevant validation recorded.
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
- Mostly clear
- Newbie friendliness
- 32/100