transformers

34e8b0a3 - Mirror the registered pytree flatten when building automatic dynamic shapes (#48578)

Mirror the registered pytree flatten when building automatic dynamic shapes (#48578) * Mirror the registered pytree flatten when building automatic dynamic shapes `get_auto_dynamic_shapes` treated every object with a `__dict__` as flattening to a flat list of tensors, which is true only for the `Cache` subclasses the exporter registers itself. A `ModelOutput` is registered by transformers with the standard dict-like flatten, so `torch.export` expects one child per field -- and a field holding a tuple keeps that tuple in the tree spec. Passing a `ModelOutput` as a generate kwarg therefore failed with `Node arity mismatch; expected 2, but got 3`. Ask the pytree registry for the flatten instead of assuming one: caches still yield a flat list, outputs keep their nesting. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * skip --------- Co-authored-by: Claude Opus 5 <noreply@anthropic.com>