gaudi_spawn

A simple launcher script for distributed training on HPUs.

Single node:

>>> python gaudi_spawn.py --world_size=NUM_CARDS_YOU_HAVE --use_mpi
           YOUR_TRAINING_SCRIPT.py (--arg1 --arg2 --arg3 and all other
           arguments of your training script)

Multi node:

>>> python gaudi_spawn.py --hostfile=PATH_TO_HOSTFILE --use_deepspeed
           YOUR_TRAINING_SCRIPT.py (--arg1 --arg2 --arg3 and all other
           arguments of your training script)

Functions

parse_args()

Helper function parsing the command line options.

Module Contents

gaudi_spawn.parse_args()[source]

Helper function parsing the command line options.

@retval ArgumentParser