Quickstart — PyTorch 2.7 documentation (original) (raw)

To launch a fault-tolerant job, run the following on all nodes.

torchrun --nnodes=NUM_NODES --nproc-per-node=TRAINERS_PER_NODE --max-restarts=NUM_ALLOWED_FAILURES --rdzv-id=JOB_ID --rdzv-backend=c10d --rdzv-endpoint=HOST_NODE_ADDR YOUR_TRAINING_SCRIPT.py (--arg1 ... train script args...)

To launch an elastic job, run the following on at least MIN_SIZE nodes and at most MAX_SIZE nodes.

torchrun --nnodes=MIN_SIZE:MAX_SIZE --nproc-per-node=TRAINERS_PER_NODE --max-restarts=NUM_ALLOWED_FAILURES_OR_MEMBERSHIP_CHANGES --rdzv-id=JOB_ID --rdzv-backend=c10d --rdzv-endpoint=HOST_NODE_ADDR YOUR_TRAINING_SCRIPT.py (--arg1 ... train script args...)

Note

TorchElastic models failures as membership changes. When a node fails, this is treated as a “scale down” event. When the failed node is replaced by the scheduler, it is a “scale up” event. Hence for both fault tolerant and elastic jobs, --max-restarts is used to control the total number of restarts before giving up, regardless of whether the restart was caused due to a failure or a scaling event.

HOST_NODE_ADDR, in form [:] (e.g. node1.example.com:29400), specifies the node and the port on which the C10d rendezvous backend should be instantiated and hosted. It can be any node in your training cluster, but ideally you should pick a node that has a high bandwidth.

Note

If no port number is specified HOST_NODE_ADDR defaults to 29400.

Note

The --standalone option can be passed to launch a single node job with a sidecar rendezvous backend. You don’t have to pass --rdzv-id,--rdzv-endpoint, and --rdzv-backend when the --standalone option is used.

Note

Learn more about writing your distributed training scripthere.

If torchrun does not meet your requirements you may use our APIs directly for more powerful customization. Start by taking a look at theelastic agent API.