Models re-download on every HuggingFace repo update

[TRGLN]

This is for bugs only

Did you already ask in the discord? Yes https://discord.com/channels/1144033039786188891/1444177940912013352/1445369297382674443

You verified that this is a bug and not a feature request or question by asking in the discord? Yes

Describe the bug

1. The toolkit redownloads models from HuggingFace even when the full model already exists locally. I already have the complete snapshot in models/..., but the toolkit still downloads everything again into C:\Users...\cache\huggingface\hub. Is this intended, or should it detect existing local weights? (It looks like the tool re-downloads all model files every time the HuggingFace repository gets any new commit — even if only the README or some tiny non-model file was changed. This happens because the global repository commit hash changes, and the system treats it as a “new version,” even though the actual weights and configs may be completely identical.)

2. Training won’t start offline even if all model files are already downloaded. When my internet was unstable, the toolkit refused to run and kept trying to connect to HuggingFace. Is it possible to allow a true offline mode if local files are present?

3. All models are forced onto the C: drive. Large models (Qwen, Z-Image Turbo) fill the system drive quickly. Can we get an option to choose a custom directory for model storage?

Attached a screenshot from wiztree showing duplicated downloads.

Thanks for all your work, really appreciate it.

[mookiexl]

You can make it use a local copy by giving it an absolute path to your local snapshot by giving it something along the lines of D:\models\stable_diffusion\mycheckpoint instead of huggingface name like stabilityai/stable_diffusion (notice the use of \ instead of / ). That way you can also keep the models on a different drive than the system one.

[TRGLN]

Thanks, I didn’t notice this setting inside the training menu — I was looking for it in the global settings. So points 2 and 3 regarding the local model path are resolved.

Point 1 can probably still be considered a bug, since re-downloading all weights because of a README commit feels excessive and unexpected. I’ll leave that for Ostris to decide.

[wideload1971]

Thanks, I didn’t notice this setting inside the training menu — I was looking for it in the global settings. So points 2 and 3 regarding the local model path are resolved.

Point 1 can probably still be considered a bug, since re-downloading all weights because of a README commit feels excessive and unexpected. I’ll leave that for Ostris to decide.

If you don't mind me asking, where did you find that setting inside the training menu? Like you, I also noticed the redownloading of models and would love to make it all local. I've looked but kind find anywhere for that info to go but I'm also not overly familiar with AI-Toolkit yet.