Nicolas Patry

> @Narsil thank you for the review! Before I go in and apply the comments, I thought it might be worth discussing the required (or not) `image` argument at a...

> would you be open to allowing the user to pass in (optional) words and boxes I have the same uneasiness with **any** optional inputs. Either the pipeline needs the...

Ok, thanks for all the explanation ! Now I think I am starting to understand it and all the use cases you displayed really make sense ! I think we...

> With all of this context in mind, do you have a preference for whether we extend the existing VisualQuestionAnsweringPipeline or isolate this logic into a DocumentQuestionAnsweringPipeline? I'm okay with...

> Figure out padding (specifically -- using "return_tensors" basically requires padding, so I could either enforce it or do the unsqueeze trick used in the qa pipeline) Not sure I...

> If I'm understanding the QA pipeline code correctly, the reason padding is relevant is that if you stride a document (e.g. one with more than 512 words), then one...

@ankrgyl Here are some tests we can integrate If you're ok (feel free to modify, the important part is to have exact values in the asserts everywhere except `run_pipeline_test`. https://github.com/huggingface/transformers/pull/18732/commits/2e8b01cd5e65aa64956e0f5e56e29ea8391c3955

> 1 - Is it possible to use Processors in pipelines? I think some (but not a whole lot) of the logic for Donut, and a whole lot of the...

> # ValueError: 'p5' is not in list The small model is just configured with much less layers, and `p5` is not expected to be there. (The goal is to...

100% agree with @SaulLu . There might be a use case, but it doesn't seem as a blatant missing feature (and we try to focus on those). Future reader, make...