mmg10/ptl_dist

PyTorch Lightning for Distributed Training

This repo contains sample code for distributed training for different configurations.

Distributed Data Parallel

• Data used: MNIST

Multi-Node, Single GPUs

• Code
• T4 GPU on 2 instances Command:

MASTER_ADDR={IP of RANK 0} MASTER_PORT=29500 NODE_RANK=0 python main.py
MASTER_ADDR={IP of RANK 0} MASTER_PORT=29500 NODE_RANK=1 python main.py

Single-Node, Multiple GPUs

Code
Command:

pip install torch_tb_profiler if profiling needed; else disable it
MASTER_ADDR=localhost MASTER_PORT=29500 WORLD_SIZE=2 NODE_RANK=0 python main.py
tensorboard --logdir=./tensorboard/ --host=0.0.0.0 # to view tensorboard

Distributed Model Parallel

• Data used: Intel Image Classification from

Strategy: FSDP

Multi-Node, Single GPUs

Code

• ViT Model is used
  Caveats:
• Number of devices = 2 (1 per node; 2 in total). This is in contrast with DDP training.
• PTL checkpointing doesn't work. The weights are not stored.
• Models are saved manually and then the best model is copied to the root folder.

export MASTER_PORT=29500
export MASTER_ADDR=172.31.10.239
export WORLD_SIZE=2
export NODE_RANK=0 # and 1 respectively

python -m torch.distributed.run \
    --nnodes=$WORLD_SIZE \
    --master_addr $MASTER_ADDR \
    --master_port $MASTER_PORT \
    --node_rank $NODE_RANK \
    main.py

Single Node, Multiple GPUs

Code

MASTER_ADDR=localhost MASTER_PORT=29500 WORLD_SIZE=1 NODE_RANK=0 python main.py