Arthur

Will review again!

Thanks a lot for your contributions! And congrats on the PR 😉

Hey, thanks for the contribution! I agree with you, wee should not save the filters as they just depend on the parameters with which they were created, which is why...

For the remaining failing test, I suggest you rebase on main 😉

You can also modify the test to make the CI go green 😉

The test works locally, merging 😉

3. An example would be the following: ```python class WhisperForConditionalGeneration: ... # redefine generate with custom kwargs like `task`, `return_timestamps` and `is_multilingual` def generate( self, inputs: Optional[torch.Tensor] = None, generation_config=...

Closing as I have not more comments! Things look good for now!

Hey, not really, I would say the least changes the better!

Also for the eos_token ids, you can just set it as an argument in the clip config, and maybe raise some kind of warning if it is not the last?...