Select GPU from cli

[DuaneNielsen]

The CLI has a flag --gpus

In a system with more than 1 GPU, is there a way to select the GPU you want to run on from CLI?

I tried --gpus [1] to select cuda:1 but it doesn't work.

Also the auto gpu selection didn't work for me. It tried to put the job on cuda:0, but cuda:0 didn't have enough memory to run it.

in the end I resorted to CUDA_VISIBLE_DEVICES, but that seems silly.

Thanks!

[carmocca]

I'm assuming you are referring to the LightningCLI

In that case, just do python yourscript.py --trainer.gpus=[1]

[tshu-w]

for those who use zsh, maybe we need \[1\]

[carmocca]

Alternatively, --trainer.gpus='[1]'

[tshu-w]

Thx, it looks nicer.

[DuaneNielsen]

Since I was using

import pytorch_lightning as pl
from argparse import ArgumentParser

parser = ArgumentParser()
parser = pl.Trainer.add_argparse_args(parser)
args = parser.parse_args()

trainer = pl.Trainer.from_argparse_args(args)

I ended up doing

my_program.py -gpus 1,

The comma caused pl to interpret the argument as a list of gpu device ids instead of the number of gpus.