Automatic Checkpoint Converter to HybridModel
Open
enhancement
- Dominant language
- Python
- Stars
- 17.9k
- Forks
- 4.5k
- Avg merge
- 4d 3h
- Merged PRs (30d)
- 272
Description
As we are migrating from `GPTModel` to `HybridModel`, it would make the process easier for users if `GPTModel` checkpoints can be converted automatically before training begins.
We will introduce a `--use-legacy-model` flag:
- HybridModel ckpt -> HybridModel training
- GPTModel ckpt + `--use-legacy-model` -> GPTModel training
- GPTModel ckpt -> HybridModel training
Contributor guide
Assessment
This issue has not been assessed yet.