Do models have to be rebuilt for Flash Attention?

[7k50]

Do the models (.bin and .mlmodelc for Metal/Core ML) have to be rebuilt to enable them with Flash Attention (--flash-attn) per Whisper.cpp v1.6.0?

If so, do both .bin and .mlmodelc have to be rebuilt?

I'm assuming that the answer can be inferred if one better understands the architecture and relationship between all these components, but I'm not versed or literate in that (my apologies).

[ggerganov]

No need to rebuild the models. Note that if you use CoreML + Flash Attention, then only the Decoder will utilize the FA kernels. The Encoder would run as usual (i.e. without FA)