How do you determine which model is allowed per device? fallbackModelSupportConfig

[saamerm]

I downloaded the TestFlight app, and noticed that the iPad 9 (iPad12,1) had a very few model options to select from, when compared to other devices I tested with. It got me to this fallbackModelSupportConfig here:
https://github.com/argmaxinc/WhisperKit/blob/9c673e35193fe4e67593dd2973c157ebe666e598/Sources/WhisperKit/Core/Models.swift#L1598

How do you determine what the maximum model size is per device?
We just needed openai_whisper-small.en but it seems like openai_whisper-base.en is the max model size?

Does the library use any of the hardware-level support for data formats like FP16/AMX or is it purely because of the size of the CPU/RAM?

[atiorh]

Thanks for the note @saamerm!

• Device support is based on chip capabilities and RAM at the same time.
• The primary rules: (1) A14 and newer can run all models and (2) iOS can run models below 1 GB.
• iPad12,1 has A13 whose Neural Engine is limited from a chip capability perspective (not RAM), hence the limited list of supported models.
• Our device support mappings only consider the pure Neural Engine deployment configuration. You could run whisper-small on iPad12,1 if you select the GPU but it would not be a recommended/tested configuration.