[hexonal]

The device <select> in the webui is authored CPU-first, so the default selection is CPU on every machine — including ones with a working GPU.

Impact

On Apple Silicon this means the demo ships a default that ignores the GPU entirely. Kronos-small inference on 60m bars is roughly an order of magnitude slower on CPU than on MPS, and nothing in the UI hints that a faster device is available — the dropdown looks like a deliberate choice rather than an accident of option ordering. The same applies to a CUDA box where the user does not think to change it.

Why detect server-side

The browser cannot see which backends torch was built against. torch.cuda.is_available() and torch.backends.mps.is_available() are the only honest answers, and they only exist in the Python process. So /api/available-models now also returns recommended_device, and the page applies it when it populates the dropdown.

Behaviour

• cuda if available, else mps if available, else cpu
• Falls back to cpu whenever torch is missing or the model library is unavailable, so the endpoint keeps working in the simulated-data mode
• Only preselects an option that actually exists in the dropdown, so it degrades to the current behaviour rather than setting an unusable value
• The helper text under the dropdown says which device was detected, so the choice is visible instead of silent

The user can still pick anything they want — this only changes which option starts selected.