diffusers

2b9edfbc - Check LoRA-guarded unfusing and split-projection restoration per model

Check LoRA-guarded unfusing and split-projection restoration per model Two gaps the previous commit left at the model level. Unfusing rewrites the fused weights the same way fusing does, so it refuses an unmerged adapter for the same reason -- but only the fuse side was covered, because moving the guard tests into the mixin dropped the unfuse case that test_attention_mixins.py used to check. It is back, against every model rather than the two it was hand-written for. The round trip also never checked what in-place fusion is for: that to_q, to_k and to_v are actually removed, and that unfusing brings them back. It now maps each fused projection to the split ones it replaces and asserts, per module, that in-place fusion removes them, a non-in-place fusion keeps them, and unfusing restores every one. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>