Demucs is slower and lower gpu utilisation than UVR with the same model?

❓ Questions

I’ve run quite a few comparison splitting jobs with both Demucs downloaded here and Ultimate Vocal Remover - both using htdemucs-ft as the model and options to use the Apple Silicon gpu.

In all tests, uvr is quite a lot quicker and looking at the cpu/gpu utilisation; uvr maxes out the gpu but demucs here never seems to go above 25% gpu utilisation.

Seems odd as I thought uvr was just a front end to the same cli tool?

Does anyone know what could be causing this and how to get demucs here to fully utilise the gpu?

Cheers