Re-authoring torch whisper model and convert it to tflite

[yangjundy]

Description of the bug:

It would be highly beneficial to have the supports for the popular Torch Whisper model. Currently, the model is implemented with a dynamic kv_cache, making it extremely challenging to convert to TFLite. Attempts to modify the model to use a fixed-size kv_cache have revealed additional difficulties, particularly due to the dynamic slicing of the kv_cache during its updates, the tflite conversion has not been successful.

Rewriting the model with ai-edge-torch as a reference would be incredibly helpful in overcoming these challenges.

Actual vs expected behavior:

The torch implementation can be converted to tflite.

Any other information you'd like to share?

No response

[pkgoogle]

Hi @yangjundy, thanks for the recommendation. We'll take a look.