Accelerate 1.15.0 — warning_once keeps adapter instances alive (measured) A Hugging Face forum reply analyzing Diffusers issue #15305 and PR #15307 reports that the correct fix for the `final_layer.adaLN_modulation.1` LoRA conversion is to reorder the output rows of `lora_B` rather than change the LoRA rank coordinates, leaving `lora_A` unchanged under the mapping P(BA) = (PB)A. The reply cites the original poster's test, in which stock `diffusers==0.43.0` produced a maximum absolute delta difference of about `39.0527` while the tested head of PR #15307 produced `0.0` with the converted `lora_A` equal to the original, though the author notes he did not independently rerun the comparison and treats those figures as the tester's measured result for that reproducer and head. The author cautions that full-weight checkpoints, non-LoRA factorizations, or converters with different tensor layouts may need their own mapping, and recommends checking that `A` matches the original and that `B_converted @ A_converted` matches `B @ A`, recording maximum absolute difference rather than a single `allclose` boolean. My mistake… The following is a reply to the thread about Diffusers: Diffusers 15305 — tested the fix in PR 15307: exact https://discuss.huggingface.co/t/diffusers-15305-tested-the-fix-in-pr-15307-exact/191952 For now, I took a quick look at a few related implementations and earlier fixes: Thanks for putting the result in the forum, and for being precise about which code you tested. Your report is especially useful because it keeps three things separate: the released diffusers==0.43.0 behavior, the behavior of the proposed fix at the PR head you tested, and the fact that lora A itself was left unchanged. Those are substantially more informative observations than a generic “the conversion works now” claim. My reading of issue 15305 https://github.com/huggingface/diffusers/issues/15305 , PR 15307 https://github.com/huggingface/diffusers/pull/15307 , and the corresponding model/converter implementations is that the output ordering is what must change, not the LoRA rank coordinates . In the affected final-layer mapping, the source convention is shift, scale and the Diffusers target convention is scale, shift . With a conventional two-factor LoRA update delta W = B @ A , a permutation of output channels belongs on the output rows of B . A projects into the rank space; its rows are not the source model’s shift and scale output channels. That gives the small but important invariant: source delta: B @ A target delta: swap output halves B @ A converted factors: swap output halves B @ A The same mapping can be written as P BA = PB A , where P swaps the two output halves. Under that specific contract, A is deliberately unchanged. This is the reason the particular result in your report is meaningful, not just the fact that a numerical comparison returned zero. As reported in your test https://discuss.huggingface.co/t/diffusers-15305-tested-the-fix-in-pr-15307-exact/191952 , stock 0.43.0 gave a maximum absolute delta difference of about 39.0527 , whereas the tested head of 15307 gave 0.0 , with the converted lora A equal to the original. I have not independently rerun that Diffusers converter comparison here, so I would treat those numbers as your measured result for that reproducer and head , supported by the algebra and the PR’s own converter-specific tests, rather than as an assertion about every checkpoint or later release. For the exact final layer.adaLN modulation.1 LoRA-A/LoRA-B issue, the default reading seems straightforward: lora A unchanged and reorder the output rows of lora B . If this conversion path also accepts a bias attached to 0.43.0 with a particular PR head. At my November 18, 2026 check, P BA = PB A argument is strong for this two-factor, output-row permutation. A full-weight checkpoint, a non-LoRA factorization, or a converter with a different tensor layout may require its own mapping. That is a scope distinction, not a contradiction of the reported reproduction. For a future reader who encounters a different result, I’d start with three cheap checks in this order. None is intended as extra homework for the original poster. A matches the original B converted @ A converted matches the B @ A . A single allclose boolean does not say which half of that contract held. Recording a maximum absolute difference as well makes failures easier to diagnose. To me, the strongest part of your report is the combination of a fixed reproducer, an explicit version comparison, the unchanged A factor, and a composed update that matches the intended output order. The existing PR tests and the earlier, separate fixes make the reasoning easy to follow. That is already a clear and reusable technical result, with no need to turn it into a general claim about all adapter formats or to ask you to run a larger model.