Should we clamp or throw for out-of-range sampling hyperparameters?

[domenic]

Right now the Chrome implementation throws if temperature is outside the [0, maxTemperature] range, and similarly for top-K.

This might pose an interop problem, or a compat problem over time. Although ideally developers would check maxTemperature before providing a temperature value, maybe we should make it easier for them.

The proposal would be to change from throwing, to clamping to the nearest valid value. The developer could detect this clamping by checking the session's temperature (and topK).

This opens up the possibility of people writing code such as

ai.languageModel.create({ temperature: Infinity })

to get the max temperature intentionally. Probably that's fine.

[tomayac]

I think it's fine to clamp. Is this something worth logging a console warning for, though? While ai.languageModel.create({ temperature: Infinity }) definitely is intentional, someone trying ai.languageModel.create({ temperature: 1.0 }) and then ai.languageModel.create({ temperature: 1.5 }) (assuming a [0.0, 1.0] interval) may expect a change to happen, thinking the model accepts [0.0, 2.0] (as per #41 (comment)) for example.