deepspeedai / deepspeedai/DeepSpeed
[REQUEST] Autocast for BF16
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 43.1k
- Forks
- 5k
- Avg merge
- 4d 15h
- Merged PRs (30d)
- 112
Description
Is your feature request related to a problem? Please describe.
Currently for FP16 training, there is an option for autocast, while casting must be done manually for BF16 training, which is something like the following,
if model_engine.bfloat16_enabled():
data_batch = data_batch.bfloat16()
Describe the solution you'd like
An option for autocasting BF16 just like the FP16 counterpart would be great.
Or, is there any particular reason for not implementing this in the first place?
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 locating the existing FP16 autocast implementation and the model_engine.bfloat16_enabled() path. Compare how FP16 autocasting is exposed and how BF16 batches are currently handled. Done means BF16 training offers an equivalent autocast option without requiring the shown manual cast, with coverage for the relevant training behavior.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python, pytorch
- Domain
- machine-learning
- Issue type
- Feature
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Stale
- Clarity
- Mostly clear
- Newbie friendliness
- 35/100