Distributed¶
Utilities to start multiple processes from a single program invocation in distributed learning
PyTorch¶
PyTorch provides a Python module/console script to launch distributed runs. Visit PyTorch’s torchrun documentation for more details.
The following environment variables available in all processes can be accessed through the library:
LOCAL_RANK
(accessible viaskrl.config.torch.local_rank
): The local rank.RANK
(accessible viaskrl.config.torch.rank
): The global rank.WORLD_SIZE
(accessible viaskrl.config.torch.world_size
): The world size (total number of workers in the job).
JAX¶
According to the JAX documentation for multi-host and multi-process environments, JAX doesn’t automatically start multiple processes from a single program invocation.
Therefore, in order to make distributed learning simpler, this library provides a module (based on the PyTorch torch.distributed.run
module) for launching multi-host and multi-process learning directly from the command line.
This module launches, in multiple processes, the same JAX Python program (Single Program, Multiple Data (SPMD) parallel computation technique) that defines the following environment variables for each process:
JAX_LOCAL_RANK
(accessible viaskrl.config.jax.local_rank
): The rank of the worker/process (e.g.: GPU) within a local worker group (e.g.: node).JAX_RANK
(accessible viaskrl.config.jax.rank
): The rank (ID number) of the worker/process (e.g.: GPU) within a worker group (e.g.: across all nodes).JAX_WORLD_SIZE
(accessible viaskrl.config.jax.world_size
): The total number of workers/process (e.g.: GPUs) in a worker group (e.g.: across all nodes).JAX_COORDINATOR_ADDR
(accessible viaskrl.config.jax.coordinator_address
): IP address where process 0 will start a JAX coordinator service.JAX_COORDINATOR_PORT
(accessible viaskrl.config.jax.coordinator_address
): Port where process 0 will start a JAX coordinator service.
Usage¶
$ python -m skrl.utils.distributed.jax --help
usage: python -m skrl.utils.distributed.jax [-h] [--nnodes NNODES]
[--nproc-per-node NPROC_PER_NODE] [--node-rank NODE_RANK]
[--coordinator-address COORDINATOR_ADDRESS] script ...
JAX Distributed Training Launcher
positional arguments:
script Training script path to be launched in parallel
script_args Arguments for the training script
options:
-h, --help show this help message and exit
--nnodes NNODES Number of nodes
--nproc-per-node NPROC_PER_NODE, --nproc_per_node NPROC_PER_NODE
Number of workers per node
--node-rank NODE_RANK, --node_rank NODE_RANK
Node rank for multi-node distributed training
--coordinator-address COORDINATOR_ADDRESS, --coordinator_address COORDINATOR_ADDRESS
IP address and port where process 0 will start a JAX service
API¶
- skrl.utils.distributed.jax.launcher.launch()¶
Main entry point for launching distributed runs