distillation with original Whisper codebase

[afsara-ben]

I am unsure how in the openai-whisper code distillation can be done (not talking about the huggingface implementation) -- for distillation to happen, gradient has to flow throughout the forward function, but i get grad_fn=None always, even when the model has been set to train(). Any script that does simple distillation using the originial whisper codebase (this repo)?