Poetry instillation failing to download all libraries

[benku2]

Python Version: 3.10.12
Poetry Version: 1.6.1
CUDA Version: release 12.2, V12.2.140
GPU: T4 GPU

It appears that installing on poetry is not correctly downloading all the libraries, despite claiming that it is. In error.txt, the loguru library appears to be missing, which we can confirm in packages.txt. However, from the small sample I grabbed in install.txt, loguru is claimed to be installed correctly by poetry. I have also attached a Google Colab notebook that runs through the same process to repeat it. Running the notebook with any of the available GPUs results in the same error.

error.txt
install.txt
packages.txt

[zeppombal]

Hi @benku2, thanks for using tower-eval.
Have you made sure to activate your environment prior to running the bash script?

For example, in the terminal:

> source <YOUR_ENV>/bin/activate
> bash run_paper_benchmark.sh

I cannot access your notebook, unfortunately.

[benku2]

Thanks for the reply. Yes, I am attempting to utilize the environment while running the benchmark, but am still getting the issue of missing libraries. The Google Colab should now be accessible.

[zeppombal]

Thanks for sharing the colab. The repo is set up correctly; this was a Colab issue related to using python venvs.

I seem to have found the fix: if you want to use the environment in the notebook, you have to prefix every command with source /content/tower-eval-env/bin/activate; because Colab resets the shell at every cell (see this).
For example,

!cd /content/tower-eval; source /content/tower-eval-env/bin/activate; bash run_paper_benchmark.sh

That said, I still got some huggingface credential errors after running the bash script, like:

OSError: You are trying to access a gated repo.
Make sure to have access to it at https://huggingface.co/meta-llama/Llama-2-70b-hf.

These are normal (you need to login to hugginface and accept the model's terms and conditions) but I'm not sure how you add credentials in a Colab notebook, because we don't work in that environment. I searching huggingface's documentaiton.

If you are looking to test the repository, I suggest running an example config that does not depend on a gated model; for example, one using our Tower models:

python tower_eval/cli.py gen-eval --config configs/examples/gen_eval.yaml