The memory requirement for finetuning

[kuoyenlo]

Hi guys:

I am trying to re-produce the finetune (stage 2) based on the instruction, such the following command: " torchrun --nproc-per-node NUM_GPU train.py --cfg-path train_configs/minigpt4_stage2_finetune.yaml "

ALTHOUGHT I could run the demo.py without OOM, run the finetune with official dataset result in OOM in my machine (colab). Just wondering anyone has relative experience.

Thanks!

[Jason-fan20]

The same question.

[thiner]

Maybe this is helpful #164