Creating onnx file from openai/whisper-large-v3

[arthurvb]

Dear All,

To save you time just some notes how I got Olive to work for openai/whisper-large-v3.

• I had to add "save_as_external_data": true in the json (e.g. whisper_cpu_int8.json) in conversion/config and in prepost/config
• In olive-env/lib/python3.10/site-packages/olive/passes/onnx/insert_beam_search.py I removed the check of the onnx model by changing the last line to return model_proto_to_olive_model(combined_model, output_model_path, config, False) . Otherwise the process stopped with The model does not have an ir_version set properly.

Hopefully this will help some of you,

Arthur

[thewh1teagle]

Hey @arthurvb
First of all thanks a lot for keeping the notes here.

Did you succeed to run tiny / medium models as well?
I tried to follow the repo but it output only empty result.
Maybe you can see if I'm missing something?
#1291

[xiaoyu-work]

closing as whisper model is supported here: https://github.com/microsoft/olive-recipes