The following doc version has a minor bug in its usage:
https://github.com/huggingface/transformers/blob/v4.50.3/docs/source/en/model_doc/paligemma.md#single-image-inference
At the time of writing, this is the default version of the docs people come across via
https://huggingface.co/docs/transformers/en/model_doc/paligemma
This has bug where it's using the tokenized input length:
print(processor.decode(output[0], skip_special_tokens=True)[inputs.input_ids.shape[1]: ])
but it should actually be the text input length:
print(processor.decode(output[0], skip_special_tokens=True)[len(prompt): ])
Found this out by cross-referencing the HF spaces example:
https://huggingface.co/spaces/big-vision/paligemma-hf/blob/d914d44/app.py#L38
Note sure if it's worth patching the existing docs, having a new minor release, or just closing this out as a note for others.
The following doc version has a minor bug in its usage:
https://github.com/huggingface/transformers/blob/v4.50.3/docs/source/en/model_doc/paligemma.md#single-image-inference
At the time of writing, this is the default version of the docs people come across via
https://huggingface.co/docs/transformers/en/model_doc/paligemma
This has bug where it's using the tokenized input length:
but it should actually be the text input length:
Found this out by cross-referencing the HF spaces example:
https://huggingface.co/spaces/big-vision/paligemma-hf/blob/d914d44/app.py#L38
Note sure if it's worth patching the existing docs, having a new minor release, or just closing this out as a note for others.