Horizontally Fused Training Array (HFTA) is a PyTorch extension library that helps machine learning and deep learning researchers and practitioners to develop horizontally fused models. Each fused model is functionally/mathematically equivalent to an array of models with similar/same operators.

Why developing horizontally fused models at all, you ask? This is because sometimes training a certain class of models can under-utilize the underlying accelerators. Such hardware under-utilization could then be greatly amplified if you train this class of models repetitively (e.g., when you tune its hyper-parameters). Fortunately, in such use cases, the models under repetitive training often have the same types of operators with the same shapes (e.g., think about what happens to the operators when you adjust the learning rate). Therefore, with HFTA, you can improve the hardware utilization by training an array of models (as a single fused model) on the same accelerator at the same time.

HFTA is device-agnostic. So far, we tested HFTA and observed significant training performance and hardware utilization improvements on NVIDIA V100, RTX6000 and A100 GPUs and Google Cloud TPU v3.

# NVIDIA GPUs:
$ pip install git+https://github.com/UofT-EcoSystem/hfta

# Google Cloud TPU v3:
$ pip install git+https://github.com/UofT-EcoSystem/hfta#egg=hfta[xla]

TODO

1. Clone the HFTA's repo.

   # Clone the repo (mainly for fetching the benchmark code)
   $ git clone https://github.com/UofT-EcoSystem/hfta

2. Run the MobileNet-V2 example without HFTA.

   # NVIDIA GPUs:
   $ python hfta/examples/mobilenet/main.py --version v2 --epochs 5 --amp --eval --dataset cifar10 --device cuda --lr 0.01
   
   # Google Cloud TPU v3:
   $ python hfta/examples/mobilenet/main.py --version v2 --epochs 5 --amp --eval --dataset cifar10 --device xla --lr 0.01
   
   # The following output is captured on V100:
   Enable cuDNN heuristics!
   Files already downloaded and verified
   Files already downloaded and verified
   Epoch 0 took 7.802547454833984 s!
   Epoch 1 took 5.990707635879517 s!
   Epoch 2 took 6.000213623046875 s!
   Epoch 3 took 6.0167365074157715 s!
   Epoch 4 took 6.071732521057129 s!
   Running validation loop ...

3. Run the same MobileNet-V2 example with HFTA, testing three learning rates on the same accelerator simultaneously.

   # NVIDIA GPUs:
   $ python hfta/examples/mobilenet/main.py --version v2 --epochs 5 --amp --eval --dataset cifar10 --device cuda --lr 0.01 0.03 0.1 --hfta
   
   # Google Cloud TPU v3:
   $ python hfta/examples/mobilenet/main.py --version v2 --epochs 5 --amp --eval --dataset cifar10 --device xla --lr 0.01 0.03 0.1 --hfta
   
   # The following output is captured on V100:
   Enable cuDNN heuristics!
   Files already downloaded and verified
   Files already downloaded and verified
   Epoch 0 took 13.595093727111816 s!
   Epoch 1 took 7.609431743621826 s!
   Epoch 2 took 7.635211229324341 s!
   Epoch 3 took 7.6383607387542725 s!
   Epoch 4 took 7.7035486698150635 s!

In the above example, ideally, the end-to-end training time for MobileNet-V2 with HFTA should be much less than three times the end-to-end training time without HFTA.

Check the colab tutorial

• Fourth Conference on Machine Learning and Systems (MLSys'21)
  • Horizontally Fused Training Array: An Effective Hardware Utilization Squeezer for Training Novel Deep Learning Models
    • Proceedings
    • Talk
    • arXiv
    • Please refer to the MLSys'21 Artifact Reproduction Guide on how to reproduce the reported experiments.

If you use HFTA in your work, please cite our MLSys'21 publication using the following BibTeX:

@inproceedings{MLSYS2021_HFTA,
 author = {Wang, Shang and Yang, Peiming and Zheng, Yuxuan and Li, Xin and Pekhimenko, Gennady},
 booktitle = {Proceedings of Machine Learning and Systems},
 editor = {A. Smola and A. Dimakis and I. Stoica},
 pages = {599--623},
 title = {Horizontally Fused Training Array: An Effective Hardware Utilization Squeezer for Training Novel Deep Learning Models},
 url = {https://proceedings.mlsys.org/paper/2021/file/a97da629b098b75c294dffdc3e463904-Paper.pdf},
 volume = {3},
 year = {2021}
}

We sincerely appreciate contributions! We are currently working on the contributor guidelines. For now, just send us a PR for review!

HFTA itself has a MIT License. When collecting the examples and benchmarks, we leverage other open-sourced projects, and we include their licenses in their corresponding directories.

HFTA is developed and maintained by Shang Wang (@wangshangsam), Peiming Yang (@ypm1999), Yuxuan (Eric) Zheng (@eric-zheng), Xin Li (@nixli).

HFTA is one of the research projects from the EcoSystem group at the University of Toronto, Department of Computer Science.