Huaggingface_hub Cannot find the requested files in the disk cache

Cannot find the requested files in the disk cache

Hi,

For some models, (e.g., falcon), the Huggingface Hub tries to automatically look for updates of files. This does not work in the TIRA sandbox, because there is no internet connection to improve reproducibility.

With Huggingface, we have two options.

Option 1:
Point to the model via the absolute path. In this case, Huggingface can not guess under which URL a model would live. The disk cache in Tira will mount it the files to their default location, e.g., for Llama-3-8b with the most recent snapshopt, the model would be mounted to:

/root/.cache/huggingface/hub/models--meta-llama--Meta-Llama-3-8B/snapshots/1460c22666392e470910ce3d44ffeb2ab7dbd4df

Option 2:
The same should be possible via the HF_HUB_OFFLINE environment variable to not contact the HF_Hub. here is the corresponding documentation: Environment variables

Does this work for you?
I will contact you via a direct message to resolve the problem.

Best regards,

Maik