deepspeedai / deepspeedai/DeepSpeed
[REQUEST] Support mixed precision calculation with ZERO
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.
I assume the current implementation is based on full fp16. However, in some settings, we find pure fp16 (although with master fp32 parameters) is not as accurate as pytorch/amp. Thus, we have to resort back to pyotrch/amp so similar packages.
Describe the solution you'd like
I'm not sure how deepspeed can implement this scenarios, but the way how we use is to manually convert the model to partial fp16 based on some experiment results (using apex)
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
The issue does not name a file, test, or entry point to inspect. Begin by locating the ZeRO mixed-precision implementation and its existing fp16 configuration, then define and test support for selectively converting model components while preserving the requested accuracy behavior.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python, pytorch
- Domain
- distributed-systems, machine-learning
- Issue type
- Feature
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 20/100