FFDetrDetector fails when last batch has size 1 (AttributeError: 'tuple' object has no attribute 'with_nms')

[Teekyy]

Bug Summary

FFDetrDetector.extract_widgets() crashes when the final batch has exactly one page and total pages > 1.
This happens due to inconsistent return types from model.predict().

Root Cause

model.predict() returns:
• List[sv.Detections] when batch size > 1
• sv.Detections when batch size == 1

However, extract_widgets() assumes predictions is always a list and calls:
results.extend(predictions)

When predictions is a single Detections object, it gets iterated over, causing tuples to be added to results. This later breaks when calling .with_nms().

Problematic Code

commonforms/commonforms/inference.py

predictions = self.model.predict(b, threshold=confidence)
if len(pages) == 1 or batch_size == 1:
    predictions = [predictions]
results.extend(predictions)

This logic checks the total number of pages, not the current batch size.
When the final batch has size 1 (e.g. len(pages) % batch_size == 1), predictions is not wrapped.

Failing Cases

Occurs when:
len(pages) % batch_size == 1 and len(pages) > 1

Examples (page count of pdf):
4, 7, 10, 13, ...

Additional Issue

prepare_form() exposes batch_size but it currently has no effect for FFDetr because it isn’t forwarded. This also makes it impossible to use batch_size=1 as a workaround:

results = detector.extract_widgets(
    pages, confidence=confidence, image_size=image_size
)

FFDNet’s extract_widgets() doesn’t accept batch_size, so forwarding should be conditional.