Feature: Consistent cache_dir to prevent duplicate downloads

[reneleonhardt]

Feature Request

• When downloading a model, the cache_dir should be predetermined and printed
• --help should print the cache_dir location

At the moment, cache_dir isn't predetermined and passed to hf_hub_download, so the model will be downloaded twice when just trying with uvx parakeet-mlx video.mp4 before installing and using it the three recommended ways.

[senstella]

Hmm, I couldn't replicate it. As far as I know, HF's hf_hub_download should not download the same files twice since their default cache directories are ~/.cache/huggingface. If possible, could you show me the command output from both uvx parakeet-mlx video.mp4 and parakeet-mlx video.mp4 (the latter after running the uvx version once)?

[reneleonhardt]

Sorry for the confusion, without huggingface printing out valuable information like the used cache_dir / storage_folder it's not easy to debug.

I can replicate it everytime:

• No ~/.cache/huggingface exists
• uv pip install parakeet-mlx inside a 3.13 venv
• Use it for the first time, ~/.cache/huggingface exists now after downloading (6 GB, 1 GB of that is the model folder)
• uvx parakeet-mlx starts redownloading the model instead of just processing the AUDIOS
• Even when providing the exact location huggingface_hub.constants.HF_HUB_CACHE resolves to programmatically:
  HF_HUB_CACHE=$HOME/.cache/huggingface/hub uvx parakeet-mlx

I even installed devpi, but then uvx behaves correctly without redownloading even with --no-cache and a new version number 🤷

[senstella]

I still couldn't get to replicate it, but I added explict control of HF's cache_dir with --cache-dir option in CLI, and cache_dir args in from_pretrained. Perhaps setting it permanently with PARAKEET_CACHE_DIR env variable would help!

[reneleonhardt]

Thank you for trying and mitigating it, will try it after it has been released.

Regarding your new documentation, please keep in mind that HF_HOME defaults to ~/.cache/huggingface and that cache_dir defaults to HF_HUB_CACHE which defaults to ~/.cache/huggingface/hub.

[senstella]

I edited the README a little to make it more clear that HF_HUB_CACHE is $HF_HOME/hub! Thank you for catching that!

[reneleonhardt]

You're welcome!

Here's some food for thought about HF_HUB_CACHE and users needing to use other locations than the default:
ml-explore/mlx-lm#109