Launchers

Functions for launching training on distributed processes.

accelerate.notebook_launcher

( function args = () num_processes = None mixed_precision = 'no' use_port = '29500' master_addr = '127.0.0.1' node_rank = 0 num_nodes = 1 rdzv_backend = 'static' rdzv_endpoint = '' rdzv_conf = None rdzv_id = 'none' max_restarts = 0 monitor_interval = 0.1 )

Parameters

function (Callable) — The training function to execute. If it accepts arguments, the first argument should be the index of the process run.
args (Tuple) — Tuple of arguments to pass to the function (it will receive *args).
num_processes (int, optional) — The number of processes to use for training. Will default to 8 in Colab/Kaggle if a TPU is available, to the number of GPUs available otherwise.
mixed_precision (str, optional, defaults to "no") — If fp16 or bf16, will use mixed precision training on multi-GPU.
use_port (str, optional, defaults to "29500") — The port to use to communicate between processes when launching a multi-GPU training.
master_addr (str, optional, defaults to "127.0.0.1") — The address to use for communication between processes.
node_rank (int, optional, defaults to 0) — The rank of the current node.
num_nodes (int, optional, defaults to 1) — The number of nodes to use for training.
rdzv_backend (str, optional, defaults to "static") — The rendezvous method to use, such as ‘static’ (the default) or ‘c10d’
rdzv_endpoint (str, optional, defaults to "") — The endpoint of the rdzv sync. storage.
rdzv_conf (Dict, optional, defaults to None) — Additional rendezvous configuration.
rdzv_id (str, optional, defaults to "none") — The unique run id of the job.
max_restarts (int, optional, defaults to 0) — The maximum amount of restarts that elastic agent will conduct on workers before failure.
monitor_interval (float, optional, defaults to 0.1) — The interval in seconds that is used by the elastic_agent as a period of monitoring workers.

Launches a training function, using several processes or multiple nodes if it’s possible in the current environment (TPU with multiple cores for instance).

To use this function absolutely zero calls to a CUDA device must be made in the notebook session before calling. If any have been made, you will need to restart the notebook and make sure no cells use any CUDA capability.

Setting ACCELERATE_DEBUG_MODE="1" in your environment will run a test before truly launching to ensure that none of those calls have been made.

Example:

# Assume this is defined in a Jupyter Notebook on an instance with two GPUs
from accelerate import notebook_launcher


def train(*args):
    # Your training function here
    ...


notebook_launcher(train, args=(arg1, arg2), num_processes=2, mixed_precision="fp16")

accelerate.debug_launcher

< source >

( function args = () num_processes = 2 )

Parameters

function (Callable) — The training function to execute.
args (Tuple) — Tuple of arguments to pass to the function (it will receive *args).
num_processes (int, optional, defaults to 2) — The number of processes to use for training.

Launches a training function using several processes on CPU for debugging purposes.

This function is provided for internal testing and debugging, but it’s not intended for real trainings. It will only use the CPU.

< > Update on GitHub

Accelerate

Launchers

accelerate.notebook_launcher

accelerate.debug_launcher