Matt

Hi @a8nova, to answer the questions: 1) You'll probably need to convert those files too - the `IdeficsVisionTransformer` in `vision.py` seems to be a core part of the model. You...

@a8nova I added the three files with `_autotranslate.py` endings! Note that there are likely to be issues (e.g. forgetting the `name` kwarg when initializing layers even though I told it...

Hi @a8nova! Firstly the error: The problem is that TF models don't let you assign to `self.layers`, because TensorFlow reserves that as a special keyword. What you should do is...

Hi @a8nova, let's see... For 1, we usually add an argument like `return_tensors` which can take values like `tf`, `pt`, etc. You can take a look at e.g. `models/sam/processing_sam.py` for...

@a8nova No, you can skip gradient checkpointing in TF ports!

Hi @a8nova, you're right - we used to have a `TFTrainer` class but it's now deprecated. We recommend just training our TF models using the Keras API like `model.fit()`, and...

Hi @a8nova, sorry for the delay! Firstly, for `return_tensors`, generally our processors handle it like this: ```python def __init__(self, return_tensors=None): self.return_tensors = return_tensors def __call__(self, return_tensors=None): if return_tensors = None:...

No probs - let me know if it recurs!

Hi @a8nova, the models aren't intended to run with `CHW` input - they should always receive `NCHW`! The `test_resize_token_embeddings` test is mostly designed to work with text models - you...

Hi @a8nova, sorry for the Christmas-related delay. Huge congratulations on getting the tiny-random model working though, that indicates you're probably quite close to getting the whole thing working! Firstly, I'd...