GPU-sanghathan

A tiny POC implementation of distributed training for sequential deep learning models. Implemented using plain Numpy & mpi4py.

Currently implements:

• Sequential models / deep MLPs, training using SGD.
• Data parallel training with interleaved communication & computation, similar to PyTorch's DistributedDataParallel.
• Pipeline parallel training:
  • Naive schedule without interleaved stages.
  • Gpipe schedule with interleaved FWD & interleaved BWD.
  • (soon) PipeDream Flush schedule with additional inter-FWD & BWD interleaving.
• Any combination of DP & PP algorithms.

Setup

python4 -m myenv venv
pip install -e .
# M1 Macs: conda install "libblas=*=*accelerate"
python data.py
pytest

Usage

# Sequential training
python train.py
# Data parallel distributed training
mpirun -n 4 python train.py --dp 4
# Pipeline parallel distributed training
mpirun -n 4 python train.py --pp 4 --schedule naive
# Data & pipeline parallel distributed training
mpirun -n 8 python train.py --dp 2 --pp 4 --schedule gpipe

Internals