API/OpenVino: support whisper_ctx_init_openvino_encoder() with external state for parallel processing ?

[jrosener]

Whisper.cpp API provides a way to allocate externally the state using init functions such as whisper_init_from_file_with_params_no_state() + whisper_init_state() instead of internally using functions like whisper_init_from_file_with_params().

For me it is very useful to use one common context (i.e. a common loaded model) for multiple parallel transcriptions which uses there own state (using functions like whisper_full_with_state()). Although it seems to be not thread-safe according to comments in whisper.h, it is still the recommend way for sharing the loaded model across multiple transcriptions (#341 (comment)). As far as I have tested (CPU and Cuda), it works well.

However this approach can not be used if using OpenVino, because the init function whisper_ctx_init_openvino_encoder() expect that the state has been already internally allocated by functions like whisper_init_from_file_with_params(). Would it be possible to add something like whisper_ctx_init_openvino_encoder_no_state() and whisper_init_openvino_state() to the main Whisper.cpp API so it will be possible to use a common context for multiple parallel transcriptions while using OpenVino ?

[jrosener]

Oh my bad, I missed that commit from last week that seems to implement exactly what I have requested : #2464
I will test and close the issue.