How should we handle when the backing resources of an AI model object are discarded?

[nathanmemmott]

While there is a PR for never discarding downloaded resources on abort, a user agent may want to free resources to save disk space. Maybe just trigger destroy() for any DestroyableModels who rely on that resource? Also there should be mitigations for having a site triggering the user agent to free resources.

[domenic]

Can you expand on this more?

What do you mean by "backing resources", in particular? Is this about disk space? Or is this about unloading the model from GPU memory, e.g., if the tab moves to the background?

If it's about the latter, then what I would envision is that we'd re-load the model when the site tried to use the API again. If we fail to do so for some reason, then we'd use the "UnknownError" escape clause.

[nathanmemmott]

This is about disk space. For example, if we delete a language pack to free up disk space, what happens to the Translators that rely on it?

[domenic]

I think starting to fail with "UnknownError" is probably best in that case. I'll try to make the spec more explicit.

[domenic]

I pushed an extra commit into #47 (namely, 7d95479) which should help make this clearer. Let me know what you think, especially if you think it'd be useful to call this out elsewhere in the spec instead of just via the indirect reference in the "the algorithm" section.

[nathanmemmott]

That works for me!

[domenic]

#47 is merged, so closing this now!