Added cache_dir argument

[cc-crisan]

Added cache_dir argument for re-using local cache when possible.

[as-suvorov]

--cache-dir is a generic name but option enabled for text hf text models only. Does it make sense to rename it to --hf-cache-dir and support other model types (text-to-image, visual-text etc)?

[Copilot]

Pull Request Overview

This pull request adds a new cache_dir argument to support re-using a local Huggingface cache during model loading.

• Updated argument parsing in the CLI to include cache_dir.
• Modified load_text_hf_pipeline, load_text_model, and load_model to accept and pass along the cache_dir argument.

Reviewed Changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
tools/who_what_benchmark/whowhatbench/wwb.py	Added cache_dir argument to the CLI and passed its value to downstream calls in main.
tools/who_what_benchmark/whowhatbench/model_loaders.py	Updated function signatures for model loaders (e.g., load_text_hf_pipeline, load_text_model, load_model) to accept cache_dir.

Comments suppressed due to low confidence (1)

tools/who_what_benchmark/whowhatbench/model_loaders.py:381

• Consider adding the cache_dir argument to the call of load_text2image_model for consistency with the text model loaders, if the HF-based text-to-image pipeline supports cache reuse.

return load_text2image_model(model_id, device, ov_options, use_hf, use_genai)

[nikita-savelyevv]

As far as I know caching of HF models is enabled by default at /home/<username>/.cache/huggingface/hub. The default model cache location can be changed with a HF_HUB_CACHE env. variable. More info here: https://huggingface.co/docs/huggingface_hub/en/guides/manage-cache .

So if the default cache location doesn't suit you for some reason, I believe you can run wwb like this:

$ HF_HUB_CACHE=<some_path> wwb ...

@cc-crisan Have you tried the approach above? If so, what are the reasons the additional argument is still required?

[cc-crisan]

As far as I know caching of HF models is enabled by default at /home/<username>/.cache/huggingface/hub. The default model cache location can be changed with a HF_HUB_CACHE env. variable. More info here: https://huggingface.co/docs/huggingface_hub/en/guides/manage-cache .

So if the default cache location doesn't suit you for some reason, I believe you can run wwb like this:

$ HF_HUB_CACHE=<some_path> wwb ...

@cc-crisan Have you tried the approach above? If so, what are the reasons the additional argument is still required?

I tried it in a different context and it did not seem to be working. But I was expecting it to work with HF_HUB_OFFLINE=true, and in that context, it wouldn't get the revision (commit_hash) from the huggingface website (default it 'main') so it failed a regex match looking for a 40-characters string. I will re-test it and let you know.

[cc-crisan]

My colleague, @rvuia, found a workaround by specifying a direct path to the cached model. That plus HF_HUB_OFFLINE set on true seem to be working. I'll close this PR as there is no need to change wwb.py.