Do OpenVino + CoreML require the GGUF?

[peardox]

PR #3151 was intended to tidy a user's file-system

Looking at the code it appears that possibly only one version of the model is strictly required

If good versions of GGUF, OpenVINO and / or mlmodelc versions are available and the user chooses to use a OpenVINO or CoreML model is the GGUF actually required?

In the worst imaginable scenario (Apple Silicon) you could have all three available but which one gets used if all three are available as models and the required code has been included?

Would an Apple Silicon Mac use only the CoreML model?

In a similar situation on non-Apple if OpenVino was available would the bin be used?

The whisper_init_... functions all expect a GGUF - is this possibly surplus to requirements if CoreML / OpenVINO are available?

i.e. would it not be possible to alter the whisper_init functions to only use one model and not at least two?

[danbev]

If good versions of GGUF, OpenVINO and / or mlmodelc versions are available and the user chooses to use a OpenVINO or CoreML model is the GGUF actually required?

One thing to keep in mind here is that the OpenVINO and CoreML models are encoder models, and that the decoder model is still required which is the reason for also needing the ggml/gguf model.

[peardox]

Ahh, so we do need at least two then (I'd started wondering about this after writing the post)

There is, however, presumably some overlap with encoder so possibly just the decoder is actually REQUIRED if OpenVINO?