Auto-dispatching wrapper that loads a SQuAD-style QA head.
Resolves to {Family}ForQuestionAnswering across BERT / RoFormer —
models that emit per-token start / end logits over the input sequence.
Notes
The forward returns an output with start_logits and end_logits
of shape (B, T) each. Inference picks the (start, end) span that
maximises start_logits[s] + end_logits[e] subject to s <= e.
Examples
>>> from lucid.models import AutoModelForQuestionAnswering
>>> model = AutoModelForQuestionAnswering.from_pretrained("bert_base_qa")
>>> type(model).__name__
'BERTForQuestionAnswering'