transformers

e2d4ac03 - Remove many output_attentions and other traced outputs on 100+ models (#43590)

Remove many output_attentions and other traced outputs on 100+ models (#43590) * first batch, let's see * propagate changes * fixup broken tests * simplify more models * hack * fix attentions * I'm bullied by the new fix-repo * this one too * fix-repo * ...fix-repo? * change up * propagate changes again * more changes * fixes * revert CLAP to the back burner * more changes * some broken clap stuff * New batch of models * something forgotten? * biiig batch + handle explicit None values for output_attentions in generic.py * make fixup * remove dummy record outputs * batch, with some difficult ones (dpt) * ugly fix * update * fix * propagate changes * some more modifs * fixup * be careful about this stuff * style * fix repo * fix more * loose booleans * missing docstring * fix import * fixing fixing * hacky kosmos2.5... * upgrade * simplifyyy * and another one * propagate many changes * upgraaade * fix * ay * fixes, many * fiiix * more more updates * might slow things down? * fix fix * fiix * fix * fiiix meeeerge confliiiicts * style * modular * modular * small update... * tests broken * remapping a bunch of kwargs to encoder_kwargs to avoid contamination * style * modular mess up * more fixes * fix more things * more fixes * style * address comments 1 * argh * fiiix * broken kosmos2 why * decorators * done * fixup docs, * last fixes? * bring back typing! * modularities * modulaaar * style * fix compile * autodocstringss * The greatest decorator upgrade the great 2026 decorator upheaval update everything fixes? style * fix modular * few fixes * fixup wrong decorators * fix gemma * fix align * revert numba changes * fixup pp v2 modular * remove unnecessary merge with defaults and return tuple + refactor kosmos models * oops * fixup repo * pass kwargs on idefics properly * new should have beens * dang copy pasting is not my strong suite * ... no comment * fixup some pretrained stuff * refactor zamba models properly + fixup gradient ckpting within merged defaults * style * several fixups - gradient ckpting - missing merge defaults - input shape stuff - probably even more * style + fixup * fix repo * interesting * small fixes * fix align and other small fixes * fix repo * this time * p1 * p2 * fix encoder got cache during generate * let's try this * fix * revert higgs * fixup models * fix * that copy is bad * new round of fixes * fixup some backbones * style * fix xlstm training inplace op * fix repo * revert the kwargs renames * wrong decorator application * style * fix defaults for vision args * fix concistency * remove the user requested workaround, always return * style * some type annotations * p1 * p2 * fix repo * readd the workaround * fix git * fix vitpose backbone * fixup * quick fixes * repo * fix * fix these * x clip was implemented so wrong ICANT * clipseg fixups * fix repo * fixup weird hidden states treatment, backbones are inherently already tested --> only need backbone mixin tests * style * fix last tests * change backbones mixin instead with a wrapper always applied to the subclasses * style * force dict * cleanup * missed this one * wow this was nasty * refactor a few models * fixups * readd backbones to common tests * revert gc changes * fix repo * explicit decorators * revert zamba * fixup the unrelated renaming in dino 3 vit * fixup some modular stuff * fixup weird zamba2 workaround * fixup dino stuff * idefics simplifications * small fixes * fix kosmos 2.5 * fixup * fix clvp, same failure as main when fixing the return dict stuff * same fix as in the other pr * fixup some ci stuff * style * revert pr to capturing --------- Co-authored-by: vasqu <antonprogamer@gmail.com> Co-authored-by: Anton Vlasjuk <73884904+vasqu@users.noreply.github.com>