CUDA OOM.

[waitfor-night]

Can the pre-trained model be replaced with a more lightweight one? Or I'd like to know approximately how much GPU memory is required to run this project.

#! /bin/bash
environs="WORLD_SIZE=3 RANK=0 LOCAL_RANK=0 LOCAL_WORLD_SIZE=3"

运行命令

run_cmd="$environs python -m torch.distributed.run
--nproc_per_node=3
mtv/sample_video.py --base ./configs/cogvideox_5b.yaml ./configs/inference.yaml
--seed $RANDOM --input-file $1 --output-dir $2"

echo ${run_cmd}
eval ${run_cmd}

echo "DONE on hostname"

[shuchenweng]

An 40GB GPU is required for inference. A lightweight model is under research.