torchrun (Elastic Launch)¶

torchrun provides a superset of the functionality as torch.distributed.launch with the following additional functionalities:

Worker failures are handled gracefully by restarting all workers.
Worker RANK and WORLD_SIZE are assigned automatically.
Number of nodes is allowed to change between minimum and maximum sizes (elasticity).

Transitioning from torch.distributed.launch to torchrun¶

torchrun supports the same arguments as torch.distributed.launch except for --use_env which is now deprecated. To migrate from torch.distributed.launch to torchrun follow these steps:

If your training script is already reading local_rank from the LOCAL_RANK environment variable. Then you need simply omit the --use_env flag, e.g.:
torch.distributed.launch

torchrun
$ python -m torch.distributed.launch --use_env train_script.py
$ torchrun train_script.py

If your training script reads local rank from a --local_rank cmd argument. Change your training script to read from the LOCAL_RANK environment variable as demonstrated by the following code snippet:

torch.distributed.launch

torchrun

import argparse
parser = argparse.ArgumentParser()
parser.add_argument("--local_rank", type=int)
args = parser.parse_args()

local_rank = args.local_rank

import os
local_rank = int(os.environ["LOCAL_RANK"])

The aformentioned changes suffice to migrate from torch.distributed.launch to torchrun. To take advantage of new features such as elasticity, fault-tolerance, and error reporting of torchrun please refer to:

Train script for more information on authoring training scripts that are torchrun compliant.
the rest of this page for more information on the features of torchrun.

Usage¶

Single-node multi-worker

>>> torchrun
    --standalone
    --nnodes=1
    --nproc_per_node=$NUM_TRAINERS
    YOUR_TRAINING_SCRIPT.py (--arg1 ... train script args...)

Fault tolerant (fixed sized number of workers, no elasticity, tolerates 3 failures):

>>> torchrun
    --nnodes=$NUM_NODES
    --nproc_per_node=$NUM_TRAINERS
    --max_restarts=3
    --rdzv_id=$JOB_ID
    --rdzv_backend=c10d
    --rdzv_endpoint=$HOST_NODE_ADDR
    YOUR_TRAINING_SCRIPT.py (--arg1 ... train script args...)

HOST_NODE_ADDR, in form <host>[:<port>] (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.

Elastic (min=1, max=4, tolerates up to 3 membership changes or failures):

>>> torchrun
    --nnodes=1:4
    --nproc_per_node=$NUM_TRAINERS
    --max_restarts=3
    --rdzv_id=$JOB_ID
    --rdzv_backend=c10d
    --rdzv_endpoint=$HOST_NODE_ADDR
    YOUR_TRAINING_SCRIPT.py (--arg1 ... train script args...)

HOST_NODE_ADDR, in form <host>[:<port>] (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 on rendezvous backend:

For multi-node training you need to specify:

--rdzv_id: A unique job id (shared by all nodes participating in the job)
--rdzv_backend: An implementation of torch.distributed.elastic.rendezvous.RendezvousHandler
--rdzv_endpoint: The endpoint where the rendezvous backend is running; usually in form host:port.

Currently c10d (recommended), etcd-v2, and etcd (legacy) rendezvous backends are supported out of the box. To use etcd-v2 or etcd, setup an etcd server with the v2 api enabled (e.g. --enable-v2).

Warning

etcd-v2 and etcd rendezvous use etcd API v2. You MUST enable the v2 API on the etcd server. Our tests use etcd v3.4.3.

Warning

For etcd-based rendezvous we recommend using etcd-v2 over etcd which is functionally equivalent, but uses a revised implementation. etcd is in maintenance mode and will be removed in a future version.

Definitions:

Node - A physical instance or a container; maps to the unit that the job manager works with.
Worker - A worker in the context of distributed training.
WorkerGroup - The set of workers that execute the same function (e.g. trainers).
LocalWorkerGroup - A subset of the workers in the worker group running on the same node.
RANK - The rank of the worker within a worker group.
WORLD_SIZE - The total number of workers in a worker group.
LOCAL_RANK - The rank of the worker within a local worker group.
LOCAL_WORLD_SIZE - The size of the local worker group.
rdzv_id - A user-defined id that uniquely identifies the worker group for a job. This id is used by each node to join as a member of a particular worker group.

rdzv_backend - The backend of the rendezvous (e.g. c10d). This is typically a strongly consistent key-value store.
rdzv_endpoint - The rendezvous backend endpoint; usually in form <host>:<port>.

A Node runs LOCAL_WORLD_SIZE workers which comprise a LocalWorkerGroup. The union of all LocalWorkerGroups in the nodes in the job comprise the WorkerGroup.

Environment Variables:

The following environment variables are made available to you in your script:

LOCAL_RANK - The local rank.
RANK - The global rank.
GROUP_RANK - The rank of the worker group. A number between 0 and max_nnodes. When running a single worker group per node, this is the rank of the node.
ROLE_RANK - The rank of the worker across all the workers that have the same role. The role of the worker is specified in the WorkerSpec.
LOCAL_WORLD_SIZE - The local world size (e.g. number of workers running locally); equals to --nproc_per_node specified on torchrun.
WORLD_SIZE - The world size (total number of workers in the job).
ROLE_WORLD_SIZE - The total number of workers that was launched with the same role specified in WorkerSpec.
MASTER_ADDR - The FQDN of the host that is running worker with rank 0; used to initialize the Torch Distributed backend.
MASTER_PORT - The port on the MASTER_ADDR that can be used to host the C10d TCP store.
TORCHELASTIC_RESTART_COUNT - The number of worker group restarts so far.
TORCHELASTIC_MAX_RESTARTS - The configured maximum number of restarts.
TORCHELASTIC_RUN_ID - Equal to the rendezvous run_id (e.g. unique job id).
PYTHON_EXEC - System executable override. If provided, the python user script will use the value of PYTHON_EXEC as executable. The sys.executable is used by default.

Deployment:

(Not needed for the C10d backend) Start the rendezvous backend server and get the endpoint (to be passed as --rdzv_endpoint to the launcher script)
Single-node multi-worker: Start the launcher on the host to start the agent process which creates and monitors a local worker group.
Multi-node multi-worker: Start the launcher with the same arguments on all the nodes participating in training.

When using a job/cluster manager the entry point command to the multi-node job should be this launcher.

Failure Modes:

Worker failure: For a training job with n workers, if k<=n workers fail all workers are stopped and restarted up to max_restarts.
Agent failure: An agent failure results in a local worker group failure. It is up to the job manager to fail the entire job (gang semantics) or attempt to replace the node. Both behaviors are supported by the agent.
Node failure: Same as agent failure.

Membership Changes:

Node departure (scale-down): The agent is notified of the departure, all existing workers are stopped, a new WorkerGroup is formed, and all workers are started with a new RANK and WORLD_SIZE.
Node arrival (scale-up): The new node is admitted to the job, all existing workers are stopped, a new WorkerGroup is formed, and all workers are started with a new RANK and WORLD_SIZE.

Important Notices:

This utility and multi-process distributed (single-node or multi-node) GPU training currently only achieves the best performance using the NCCL distributed backend. Thus NCCL backend is the recommended backend to use for GPU training.
The environment variables necessary to initialize a Torch process group are provided to you by this module, no need for you to pass RANK manually. To initialize a process group in your training script, simply run:

>>> import torch.distributed as dist
>>> dist.init_process_group(backend="gloo|nccl")

In your training program, you can either use regular distributed functions or use torch.nn.parallel.DistributedDataParallel() module. If your training program uses GPUs for training and you would like to use torch.nn.parallel.DistributedDataParallel() module, here is how to configure it.

local_rank = int(os.environ["LOCAL_RANK"])
model = torch.nn.parallel.DistributedDataParallel(model,
                                                  device_ids=[local_rank],
                                                  output_device=local_rank)

Please ensure that device_ids argument is set to be the only GPU device id that your code will be operating on. This is generally the local rank of the process. In other words, the device_ids needs to be [int(os.environ("LOCAL_RANK"))], and output_device needs to be int(os.environ("LOCAL_RANK")) in order to use this utility

On failures or membership changes ALL surviving workers are killed immediately. Make sure to checkpoint your progress. The frequency of checkpoints should depend on your job’s tolerance for lost work.
This module only supports homogeneous LOCAL_WORLD_SIZE. That is, it is assumed that all nodes run the same number of local workers (per role).
RANK is NOT stable. Between restarts, the local workers on a node can be assgined a different range of ranks than before. NEVER hard code any assumptions about the stable-ness of ranks or some correlation between RANK and LOCAL_RANK.
When using elasticity (min_size!=max_size) DO NOT hard code assumptions about WORLD_SIZE as the world size can change as nodes are allowed to leave and join.
It is recommended for your script to have the following structure:

def main():
  load_checkpoint(checkpoint_path)
  initialize()
  train()

def train():
  for batch in iter(dataset):
    train_step(batch)

    if should_checkpoint:
      save_checkpoint(checkpoint_path)

(Recommended) On worker errors, this tool will summarize the details of the error (e.g. time, rank, host, pid, traceback, etc). On each node, the first error (by timestamp) is heuristically reported as the “Root Cause” error. To get tracebacks as part of this error summary print out, you must decorate your main entrypoint function in your training script as shown in the example below. If not decorated, then the summary will not include the traceback of the exception and will only contain the exitcode. For details on torchelastic error handling see: https://pytorch.org/docs/stable/elastic/errors.html

from torch.distributed.elastic.multiprocessing.errors import record

@record
def main():
    # do train
    pass

if __name__ == "__main__":
    main()

torchrun (Elastic Launch)¶

Transitioning from torch.distributed.launch to torchrun¶

Usage¶

Docs

Tutorials

Resources