Fix flaky RfDetr test_save_load (#48869)
* Fix flaky test_save_load for RfDetr by marking with @is_flaky
Using attention_interface (via ALL_ATTENTION_FUNCTIONS.get_interface) vs calling
eager_attention_forward directly produces a tiny floating-point difference (~1e-6)
between a freshly-initialized model and one loaded via from_pretrained. This diff
can cause top_k to select different proposal indices, making later output values
differ at scale ~1.0.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Update PR number in is_flaky description
---------
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>
