[Pfannkuchensack]

Summary

Real Qwen-Image LyCORIS files did not work:

• lycoris_transformer_blocks_… (LoKr, 43 files in 6 Hugging Face repos) installed as an unknown model.
• lora_transformer_transformer_blocks_… (ai-toolkit LoHa) installed as a FLUX LoRA.
• LoKr with dotted keys failed to load with Unsupported lora format, because grouping cut inside the module path.
• LoHa and factorized LoKr in the plain and Kohya layouts were not identified.

Now:

• Both prefixes are recognized, and the conversion handles them. They count only with a key on a Qwen-only module (img_mod, txt_mod, img_mlp, txt_mlp), so FLUX files with the same spelling stay FLUX.
• flux_format_from_state_dict no longer claims a file with such a module.
• Dotted LoKr and LoHa keys group by module path.
• A learned LyCORIS scalar is applied. It is folded into the layer before key normalization, for LoKr, LoHa and plain LoRA.

Related Issues / Discussions

Closes #10020

#10184 touches the same converter, constants, test file and qwen-image.mdx. The two do not depend on each other, but whichever lands second needs a rebase.

QA Instructions

Checks run:

• uv run --no-sync pytest -n 4 tests/backend/patches tests/backend/model_manager/configs tests/model_identification: 1949 passed, 25 skipped (CPU).
• Ruff check and format: passed.

Real files:

• Two real lycoris_transformer_blocks_ LoKr files identify as Qwen-Image and convert to 720 layers.
• One of them (mikaelh's Minecraft LoKr) installed and generated through the server on an RTX 4090. At a fixed seed, weight 1 against weight 0 gives 12.93 dB PSNR, so it takes effect.
• FLUX LoHa fixtures still identify as FLUX.

To reproduce: install a Qwen-Image LoKr whose keys start with lycoris_transformer_blocks_. It should install as Qwen-Image and change the image at weight 1.

Not run:

• Family E2E for Qwen-Image.
• Linux and ROCm. The change has no platform-specific code.

Review

Resolved during review:

• The first version of the new suffix rule claimed FLUX LoHa files for Qwen-Image. New prefixes and LyCORIS-only suffixes now need a key on a Qwen-only module, and base suffixes behave as before.
• scalar was only handled for LoKr and LoHa; it now also applies to plain LoRA layers.
• LoHa and LoKr files that carry dora_scale, such as ThatDustyGuy/QwenImage-PersonalLoRA, were identified as Qwen-Image but failed at generation with KeyError: 'lora_up.weight'. With fix(lora): handle DoRA magnitudes on LoKr and LoHa layers #10075 merged into this branch, the shared layer factory applies them as plain LoHa and LoKr without the magnitudes. The conversion tests now keep dora_scale in the ai-toolkit LoHa header and cover dotted LoHa and LoKr with it; with DoRA dispatched first again, all three fail. The family page says the magnitudes are skipped.

Remaining limitations:

• A file in these layouts that trains only attention layers cannot be told apart from other families: a lycoris_ file stays unknown, and a lora_transformer_transformer_blocks_ file stays FLUX.

Compatibility / Rollout

No migration. Identification only runs at install, so a file already installed as FLUX or unknown keeps that record until it is reinstalled.

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Meaningful regression coverage added / updated where needed; obsolete tests/code removed
• Persisted-state and API changes include required migrations / compatibility validation
• Relevant performance/efficiency opportunities considered; material claims have evidence
• Material review findings resolved and relevant checks rerun
• Documentation added / updated (if applicable)
• Updated What's New copy (if doing a release after this PR)

---

^{Open in [code]smith · Need help on this PR? Tag @codesmith-bot with what you need.}