Automatic weights download

[michal1000w]

What does this change?

This change automatically downloads weights of the model if they are not present in the checkpoints folder, removing the need of manual download and setup

How was it tested?

Checked all synthetic tests and manual tests passed

Checklist

• uv run pytest passes
• uv run ruff check passes
• uv run ruff format --check passes

[JamesNyeVRGuy]

Generally a good idea, resolve your conflicts and pass the lint and I will read again

[michal1000w]

Generally a good idea, resolve your conflicts and pass the lint and I will read again

Now it should be rebased to the new main. All tests seems to pass now

[nikopueringer]

@JamesNyeVRGuy What do you think? Shall I squash and merge?

[JamesNyeVRGuy]

Nice work — the core design is solid and the test coverage is thorough. A few things to address before merge:

Suggestions:

• The branching in create_engine (backend.py lines ~308-325) is subtle — in the auto/None path, assets are downloaded before resolve_backend so that _auto_detect_backend can find the MLX weights and choose MLX. In the explicit path, resolve_backend runs first. This ordering matters but isn't obvious; a short comment explaining why would help future readers.
• The .to(model_precision) change in inference_engine.py looks like a separate bug fix. Worth calling out in the PR description or splitting into its own commit so it's easier to bisect later.

Minor observation (non-blocking):

• _extract_path_from_output matches any line whose path ends in .safetensors — could pick up a stray log line. Probably fine in practice but worth being aware of.