transformers

2b296305 - 🚨[wav2vec2] Support attn_implementation=sdpa dispatch (#46196)

🚨[wav2vec2] Support attn_implementation=sdpa dispatch (#46196) * [SeamlessM4T/v2] Support attn_implementation=sdpa dispatch * simplify code * Init * config * prepare relative bias function * TP friendly * func comments * parameter scale * Fix multi-device error * Code quality * Refactor * Apply a refactor to wav2vec2 * Fix copy consistency * Doc * Skip flash sdpa UT * Update * rename func * scale->scaling * Refine * copy and code quality * copy and ruff * Fix CI * Fix CI * Refine data2vec_audio * Refine * refine * Fix CI * Fix `is_causal` SymBool leak in SDPA under torch.export Gate the query-length check behind the always-concrete conditions so a non-causal module or a provided mask short-circuits to a plain `False` before touching `query.shape[2]`, which can be a data-dependent `SymBool` under torch.export (e.g. sub-sampled audio in the Conformer encoder) that SDPA rejects. When causality is still possible, `query.shape[2] > 1` stays the first `and` operand so backed symints are still specialized to a bool. * Fix SymBool error and conflicts * comments * refine * ruff * refine * Add TODO * refine TODO comments * Skip dynamic export UT for SeamlessM4T/v2 speech models --------- Co-authored-by: Ilyas Moutawwakil <57442720+IlyasMoutawwakil@users.noreply.github.com>