WhisperKit init refactor

jkrukowski commented 8 months ago

added WhisperMLModel conformance, no more conditional casting
setupModels -> static setupModelFolder which makes modelFolder non optional
removed download param, I think it's not needed. When the modelFolder param is nil we download, otherwise we load the model from the directory

Possible future improvements:

load models (featureExtractor, audioEncoder and textDecoder) in parallel?
we could add forceDownload param which, in case modelFolder is provided will redownload

xiangliangX commented 7 months ago

Can the model now be used completely offline? It seems that loadTokenizer still needs to connect to hf (huggingface) to download.

jkrukowski commented 7 months ago

I'd like to avoid requiring WhisperMLModel for these because we will soon be working with MLX models that will follow the same TextDecoding protocols, but won't have associated MLModels. Everything else looks good

not much is left in this PR after this is removed so I'll close it for now

argmaxinc / WhisperKit

WhisperKit init refactor #84