How do I run the ML tests without downloading a checkpoint?

[chiruu12]

Most of the ML suite runs against synthetic checkpoints and needs no download:

cd sdk && uv sync --all-extras --dev
make test-ml

Cases that need real weights skip cleanly. If you want those too, point UNPLUG_MODEL_PATH at a local checkpoint directory or run unplug-models download tiny, then make test-ml-harness.

If you are seeing 6 errors in test_ml_integration.py, you synced without the ml extra. uv sync --all-extras --dev fixes it.