mostafaelhoushi / olla

Optimizing the Lifetime and Location of Arrays to Reduce the Memory Usage of Neural Networks

OLLA

OLLA (Optimizing the Lifetime and Location of Arrays) enables training larger deep neural networks on existing hardware. It accomplishes this with a few techniques:

• Operator order optimization — reodering tensor operators to reduce peak memory usage
• Fragmentation reduction — dynamic memory profiling and scheduling to better-utilize memory.

Our approach is described in detail on the OLLA arXiv paper. See citing below to attribute the work.

Quickstart

Installing OLLA in your Python environment is simple:

git clone https://github.com/facebookresearch/olla
pip install . [--extra-index-url <url>]

Note:

• The above install will attempt to install torch, torchaudio, torchvision, and torchtext based on default distributions. To install for your CUDA version/OS, see the PyTorch Getting Started documentation, appending the --extra-index-url flag and value to the above install command as needed.
• OLLA is tested with Gurobi 9.1.1; use your own license or version as needed.

Benchmarks

To run benchmarks:

python benchmarks.py

Running Tests

Run all unit tests with:

python -m unittest discover -s tests --pattern "*_test.py"

Run unit tests that are skipped with by setting RUN_SKIPPED=1 in the environment before the command.

Citation

If you use OLLA, please use the below BibTex for citing:

@article{steiner2022olla,
  title={OLLA: Optimizing the Lifetime and Location of Arrays to Reduce the Memory Usage of Neural Networks},
  author={Steiner, Benoit and Elhoushi, Mostafa and Kahn, Jacob, and Hegarty, James},
  doi = {10.48550/arXiv.2210.12924},
  year={2022},
}