TorchElastic

TorchElastic allows you to launch distributed PyTorch jobs in a fault-tolerant and elastic manner. For the latest documentation, please refer to our website.

Requirements

torchelastic requires

• python3 (3.8+)
• torch
• etcd

Installation

pip install torchelastic

Quickstart

Fault-tolerant on 4 nodes, 8 trainers/node, total 4 * 8 = 32 trainers. Run the following on all nodes.

python -m torchelastic.distributed.launch
            --nnodes=4
            --nproc_per_node=8
            --rdzv_id=JOB_ID
            --rdzv_backend=etcd
            --rdzv_endpoint=ETCD_HOST:ETCD_PORT
            YOUR_TRAINING_SCRIPT.py (--arg1 ... train script args...)

Elastic on 1 ~ 4 nodes, 8 trainers/node, total 8 ~ 32 trainers. Job starts as soon as 1 node is healthy, you may add up to 4 nodes.

python -m torchelastic.distributed.launch
            --nnodes=1:4
            --nproc_per_node=8
            --rdzv_id=JOB_ID
            --rdzv_backend=etcd
            --rdzv_endpoint=ETCD_HOST:ETCD_PORT
            YOUR_TRAINING_SCRIPT.py (--arg1 ... train script args...)

Contributing

We welcome PRs. See the CONTRIBUTING file.

License

torchelastic is BSD licensed, as found in the LICENSE file.