declare-lab / instruct-eval

This repository contains code to quantitatively evaluate instruction-tuned models such as Alpaca and Flan-T5 on held-out tasks.
https://declare-lab.github.io/instruct-eval/
Apache License 2.0
528 stars 42 forks source link

Evaluate on a single 24GB/32GB GPU #29

Open lemyx opened 9 months ago

lemyx commented 9 months ago

Hi, on a single 4090 GPU with 24GB memory, the following command will cause out-of-memory.

python main.py mmlu --model_name llama --model_path huggyllama/llama-7b

After that, I try executing the command on A100-40GB, the nvidia-smi result is

image

It seems that neither 4090/3090 with 24GB memory or V100 with 32GB memory cannot test Llama-7B on mmlu under above command.

So how to evaluate llama-7b on mmlu on 24GB or 32GB GPU? any more options to enable?

Thanks

lemyx commented 9 months ago

It seems that the CUDA memory will increase during execution of the script

image

Maybe related to maximum sequence length

image

Finally, the inference can be finished on a single A100-40GB card

image