[Shanmukh-l33t]

Summary

SimpleTuner's LyCORIS training writes LoRA/LoKR/LoHA layers named by flattening the model's own diffusers module
path with underscores behind a lycoris_ prefix (lycoris_single_transformer_blocks_0_attn_to_k.lokr_w1), or
behind lycoris__. No probe recognized the layout, so the files installed as unknown on every base and forcing
the base failed in the converters (#10174's table). They now identify and convert on FLUX.1 (both spellings),
FLUX.2 Klein, SDXL, Z-Image and Wan, the way SD 3's own flat spelling already did.

• A new lycoris_flat_lora_constants resolves each flattened name against the family's module list (module names
  contain underscores of their own, so only the whole list can unflatten them). A family is claimed on names only
  it holds — single_transformer_blocks.N.attn.to_q for FLUX.1, to_qkv_mlp_proj / ff.linear_in for FLUX.2
  Klein, the UNet's deep transformer_blocks for SDXL, layers.N.attention for Z-Image, blocks.N.attn1 for
  Wan — with the width check that keeps Z-Image off Ideogram 4 and SDXL's block depth that keeps it off SD 1/2.
  Names several families share without one only-this-family name identify nothing: SD 3's flat files stay SD 3
  (locked in by test_the_sd3_flat_layout_is_still_sd3), files of no family stay unknown.
• The conversion maps the names back to diffusers paths and reuses each family's existing route: FLUX.1 through
  the diffusers-to-BFL mapping (fused QKV, swapped final modulation) unchanged, Wan through the kohya converter's
  lora_unet_blocks_<idx>_<tail> names, SDXL through the SD converter's flattened lora_unet_ keys. A key that
  names no module of the claimed family raises instead of being dropped.
• Wan's variant detection reads the inner dimension from LoKR factors the way it reads a low-rank pair, so the
  issue's 5120-wide Wan file records its variant.

Before → after on the issue's files (key layouts captured from their safetensors headers):

Base	File	before	after
FLUX.1	toxicwind/glassobject-lokr-flux	unknown	FLUX.1, converts
FLUX.1 (lycoris__)	jimmycarter/flux-training-losercity-next-lycoris5	unknown	FLUX.1, converts
FLUX.2 Klein	gattaplayer/besch-flux2-klein-9b-lokr-lion-5e-6-bs2-ga2-v01	unknown	FLUX.2 Klein, converts
SDXL	davidrd123/Hiroshige-SDXL-LoKr	unknown (Unrecognized SDXL LoRA key prefix when forced)	SDXL, converts
Z-Image	noodlecake/Retro_Anime_Painting_lokr-R8-LR2-HIGHNOISEBIAS	unknown	Z-Image, converts
Wan	uxoah/wan2.1_lokr_simpletuner	unknown	Wan (variant A14B), converts

Related Issues / Discussions

Fixes #10174.

The issue's Qwen-Image row is covered by #10189 (its probe takes Qwen-Image's own module names as evidence), and
the SD 3 rows already worked. #10184 and #10189 touch adjacent converter and test files; this PR touches only
configs/lora.py in common with them, in different functions.

QA Instructions

Regression first: tests/backend/model_manager/configs/test_lycoris_flat_lora_config.py fails on unmodified main
showing exactly the misdetection — assert [] == ['LoRA_LyCORIS_FLUX_Config'] (and the same for Flux2, SDXL,
Z-Image, Wan) in test_a_flat_lycoris_lora_is_claimed_by_its_family_alone[*] and
test_a_flat_lycoris_lora_records_its_base[*]: 12 failed, 2 passed (the two negative controls pass on main too).

Checks run (CPU, .venv/bin/python -m pytest in a fresh uv sync --extra test environment):

• pytest -n 4 tests/backend/patches: 616 passed, 24 skipped.
• pytest -n 4 tests/backend/model_manager/configs: 1578 passed, 1 skipped, 13 failed — the 13 are all in
  test_model_path_validation.py and fail identically on unmodified main in this environment (verified by
  stashing this branch's changes and re-running the file: the same 13 failures), unrelated to LoRA code.
• pytest -n 4 tests/model_identification: 129 passed.
• pytest -n 4 tests/backend/model_manager/load tests/test_model_manager: 1301 passed, 127 skipped (the loader
  gained the new dispatch).
• ruff check and ruff format on every changed path: clean.

Not run: the full test suite (large) and the frontend checks (no frontend changes).

The real-file evidence is at the key-layout level: the new fixtures are the six issue files' safetensors headers,
and every converted layer is resolved against the model it patches — Invoke's BFL FLUX class for FLUX.1, and
diffusers' own models (Flux2Transformer2DModel, ZImageTransformer2DModel, WanTransformer3DModel, and an SDXL UNet
with the released block structure) for the rest — with the FLUX.1 layers' weight deltas checked against the real
modules' shapes. Applying one through a live generation was not done (no GPU here).

Review

No material review findings; this is the initial version. Known limitations, in the same spirit as the existing
probes' documented ones:

• A FLUX.1/FLUX.2 file that trains only module names SD 3 shares (transformer_blocks.N.attn.to_q, ...) carries
  no only-this-family name and stays unidentified rather than being guessed at.
• An SDXL file confined to transformer_blocks indices 0–1 is not distinguished from SD 1/2 — the limitation
  _state_dict_looks_like_sdxl_unet_lora already documents.
• FLUX.3's flattened lycoris_… names are a different layout and remain unrecognized, as its family page says.

Compatibility / Rollout

No migrations and no API or persisted-state changes. Identification runs at install, so a file already installed
as unknown keeps that record until it is reinstalled (Re-identify model in the Model Manager).

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.}