New to ODB? Start with the main package, then open the example closest to your training stack.
| Entry | Use it for |
|---|---|
| Main package | Install ODB, read the API, and choose an integration path |
| PyPI package | Install the released package |
| Quickstart | Try the minimal PyTorch loop and understand the ODB-ready batch contract |
| Example projects | Run a framework-specific example workflow |
The organization is split into a core package and small framework-specific examples.
| Repository | Role |
|---|---|
online-dynamic-batching |
Core package, integration guides, API docs, tests, and benchmark notes |
odb-mm-mix-example |
Shared public MM-Mix-style data recipe and local TMDB utilities |
odb-example-llamafactory |
LLaMA-Factory training and evaluation workflow |
odb-example-hf-trainer |
Hugging Face Trainer workflow |
odb-example-accelerate |
Accelerate custom-loop workflow |
odb-example-lightning |
Lightning Trainer workflow |
ODB is introduced in Online Dynamic Batching with Formal Guarantees for LLM Training by Dian Li, Zekun Wang, Yaoru Wang, and Jiahong Yan.
- Observe real runtime lengths instead of relying on stale offline caches.
- Keep batching at the DataLoader/collate boundary.
- Make distributed dynamic batching explicit with aligned grouping metadata.
- Keep examples framework-specific so users can copy the path closest to their stack.
Apache-2.0.