deepspeedai / deepspeedai/DeepSpeed
Some question of gradient accumulation
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 43.1k
- Forks
- 5k
- Avg merge
- 4d 15h
- Merged PRs (30d)
- 112
Description
engine.backward(loss)
engine.step()
Question 1: When using DeepSpeed, before calling engine.backward(loss), do I need to manually divide the loss by the gradient accumulation steps, or does DeepSpeed handle this scaling internally during backward propagation?
Question 2: At the end of an epoch, if the remaining batch does not reach the full gradient accumulation steps (e.g., only 2 steps instead of the configured 4), will DeepSpeed’s model.step() automatically handle this incomplete accumulation, or do I need to manually trigger gradient updates?
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 by tracing the DeepSpeed training entry points represented by engine.backward(loss) and engine.step(). Check how configured gradient accumulation affects loss scaling and an incomplete final accumulation, then document the confirmed behavior and required caller actions for both questions.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- distributed-systems, machine-learning
- Issue type
- Documentation
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 25/100