Distributed Runner

DistributedRunner

class optimum.habana.distributed.DistributedRunner

<source>

( command_list: typing.List = [] world_size: int = 1 hostfile: typing.Union[str, pathlib.Path] = None use_mpi: bool = False use_deepspeed: bool = False use_env: bool = False map_by: bool = 'socket' multi_hls = None )

Set up training/inference hardware configurations and run distributed commands.

create_multi_node_setup

<source>

( )

Multi-node configuration setup for DeepSpeed.

create_single_card_setup

<source>

( use_deepspeed = False )

Single-card setup.

create_single_node_setup

<source>

( )

Single-node multi-card configuration setup.

create_single_node_setup_deepspeed

<source>

( )

Single-node multi-card configuration setup for DeepSpeed.

create_single_node_setup_mpirun

<source>

( )

Single-node multi-card configuration setup for mpirun.

process_hostfile

<source>

( ) → str

Returns

str

address of the master node.

Returns the master address to use for multi-node runs with DeepSpeed. Directly inspired from https://github.com/microsoft/DeepSpeed/blob/316c4a43e0802a979951ee17f735daf77ea9780f/deepspeed/autotuning/utils.py#L145.

run

<source>

( )

Runs the desired command with configuration specified by the user.