Question about Training Time

[Backdrop9019]

Hello,

Thank you for sharing your excellent research and code. I am currently pretraining an image encoder using 8 A100 GPUs. The estimated time of arrival (ETA) is about 6 hours. Is this normal? Could you share the pretraining and fine-tuning times along with the number of GPUs used for each setting? It would be very helpful.

Thank you!

[Backdrop9019]

I also want to reproduce using vicuna-7b. Do I only need to modify the script to "BASE_LLM_PATH=lmsys/vicuna-7b-v1.5"? Or are there additional settings or hyperparameters that need to be changed?