LightBinPack is a lightweight library for solving bin packing problems, implementing core algorithms in C++ and providing a Python interface, and has been optimized for bin packing algorithms in both pre-training and post-training scenarios for LLMs.
pip install lightbinpackIf you want to use the latest features, you can install the development version from GitHub:
pip install git+https://github.com/TechxGenus/LightBinPack.gitHere is a simple example:
from lightbinpack import pack
lengths = [20, 20, 10, 10, 10, 10]
batch_max_length = 40
# For linear attention or disabled document mask
results = pack(lengths, batch_max_length, variant="linear")
print(results)
# For attention with document mask enabled
results = pack(lengths, batch_max_length, variant="square", dp_size=2)
print(results)The library also includes optimizations for heterogeneous nodes, prefix sharing, and context parallelism scenarios. You can find more examples and documentation in the docs and examples directories.
- Multipack Sampler
- OBFD
- Claude and GPT(o1)
Contributions are welcome! If you find any bugs or have suggestions for improvements, please open an issue or submit a pull request.