- Add module_skip_keys entry for HiDreamO1Qwen3VLTransformer so SDNQ
leaves embedding/output layers (lm_head, embed_tokens, x_embedder,
t_embedder1, final_layer2, patch_embed, pos_embed) at full precision.
- Flip allow_quant to True in load_hidream_o1 and surface the resolved
quant type in the load log.
`get_version()` previously stored the raw `git remote get-url origin`
output in `ver['url']`. For ssh-form remotes (`git@host:owner/repo` or
`ssh://git@host/owner/repo`), the GitHub branches API url builder in
`check_version()`, which strips the `https://github.com/` prefix,
produced a malformed url that returned 404. The 404 response (a dict
with `message`) was then iterated by a list comprehension filtering on
`'name' in b`, yielding an empty branch list and a misleading
"branch=dev skipping update" warning.
- normalize ssh and ssh:// remotes to canonical https form in
`get_version()`, so every consumer of `ver['url']` sees one shape
- guard `check_version()` against non-list api responses with an
`isinstance` check that logs the actual payload instead of silently
yielding `[]`
Rename `autocomplete_replace_underscores` to `autocomplete_keep_underscores`
(default false) so both autocomplete toggles read as "keep" semantics.
`insertTag` now honors the setting for both tag and artist modes, matching
the dropdown display. Embedding insertion always preserves underscores
because the names are filesystem identifiers.
- ui_definitions: rename option, flip default, update label
- scripts/autocomplete: rename handler, checkbox, and config-bridge keys
- javascript/autocomplete: invert render polarity; insertTag takes a `kind`
parameter; drop the artist-mode hardcoded underscore replacement so the
setting controls tags and artists uniformly
Closes#4827
Move the family-iteration loop into pipelines/flux/flux2_lora.try_load
so the f2 branch in lora_load.load_safetensors becomes a one-line call.
Per-family try_load_* entry points stay public for direct testing.
Addresses review feedback on PR #4841.
Gate finds the on-disk preview via resolve_preview_file and a
non-empty UserComment check so previews lacking embedded params enter
the worker. embed_preview_parameters writes via temp + os.replace,
converts RGBA to RGB before JPEG save, removes unreadable sources, and
invalidates any co-located .thumb.jpg. get_exif decodes EXIF
UserComment so regenerated thumbs propagate the source's params. For
.mp4 and .webm previews, embed targets the .thumb.jpg frame extracted
by save_video_frame instead of the unembeddable video file.
backfill_preview_parameters and resolve_preview_file share
find_ui_preview_file, which mirrors find_preview's iteration order so
backfill embeds into the file the modernUI actually surfaces (handles
LoRAs with multiple coexisting variants like .jpg + .jpeg).
civitai_meta_to_parameters drops generator-specific noise keys (comfy,
comfyui, workflow, extraMetadata) and applies a 512-char cap on
pass-through values to keep the embedded chunk compact.
When download_civit_preview short-circuits at 304 (file exists), the
rescan path now falls back to backfill_preview_parameters: checks
whether the file already carries an embedded parameters chunk and, if
not, embeds img.meta in place. Existing libraries get retroactive
metadata via the existing "Search metadata" rescan, no model
re-download required.
Civitai's CDN strips PNG tEXt/EXIF chunks during delivery, leaving the
modernUI Preview metadata tab empty for downloaded networks. Format the
version API's images[].meta into sdnext's standard parameters channel
(PNG tEXt / JPEG-WEBP EXIF UserComment) so read_info_from_image picks
it up through the existing path. Embedding is wrapped in try/except
and never blocks the download.
get_method now returns (method, reason). The reason distinguishes
user opt-in (lora_force_diffusers), class-forced, hash-forced, and
arch-unsupported routes from the default native path; surfaced in
the info-level "Network load: type=LoRA load=<method>(<reason>)"
line so users can tell why a given load took a particular path.
36 tests across parsing primitives, family loaders, calc_updown shape
sanity, and the lora_apply ex_bias accumulation path. Builds a mock
Flux2-shaped transformer (HIDDEN=128, 2 double + 2 single blocks) and
writes synthesized state dicts to tempfile-managed safetensors so the
loaders' real read_state_dict path runs.
Loader fixtures include the full nine-family menu plus four prefix
variants (kohya, BFL/AI-toolkit, PEFT, bare-BFL) and the two real
ground-truth formats observed on Hugging Face: base_model.model.
(peft.save_pretrained wrapper) and lycoris_ (LyCORIS-standalone).
Discriminated from OFT by 4-D oft_blocks shape (boft_m, block_num,
block_size, block_size), mirroring LyCORIS algo_check. The cascade
of log_2(block_num)+1 Cayley rotations is applied via the
unflatten/transpose/flatten reshape sequence in NetworkModuleBOFT,
porting LyCORIS make_weight verbatim.
base_model.model. is the peft.save_pretrained wrapper; stripped
via _unwrap_peft_wrapper and re-parsed against the standard prefix
list (so the same code path handles whatever format peft wrapped).
lycoris_ is LyCORIS-standalone (e.g. SimpleTuner LoKR); base path
is an underscore-rendered diffusers path returned verbatim.
Adds 'f2' to allow_native and dispatches all eight native loaders
through load_safetensors. lora_force_diffusers continues to route
through lora_diffusers.load_diffusers as the explicit PEFT escape
hatch.
Targets the elementwise-affine RMSNorms inside attention (norm_q,
norm_k, norm_added_q, norm_added_k). lora_convert.py:502 skips
network_layer_name on transformer norm modules outside SD3, so this
loader stamps it directly on the modules it binds — surgical, no
shared-code edit, no impact on modules a Norm adapter doesn't target.
Loads the four ab.weight tensors plus alpha. Fused QKV is skipped:
GLoRA's target-dependent term doesn't row-slice cleanly without
recomputing against a fused proxy weight.
The .on_input marker disambiguates IA3 from other families that also
have .weight keys; the per-group gate insists on both. Fused QKV is
skipped with a warning.
Both kohya (oft_blocks + alpha-as-constraint) and LyCORIS (oft_diag)
layouts route through NetworkModuleOFT. Fused QKV in double_blocks is
skipped: the block-diagonal rotation's block_size = out_features /
num_blocks would need re-derivation per chunk to fit a Q/K/V split.
Standard non-Tucker LoHAs route through NetworkModuleHadaChunk for
fused QKV in double_blocks. Tucker (CP-decomposed) variants on fused
targets are skipped with a warning. Single-block linear1 and non-QKV
double-block targets work for any LoHA layout.
Adds per-family SUFFIX/MARKER tables and shared parse_key /
group_by_suffixes / resolve_targets helpers, mirroring the pattern
in pipelines/z_image/zimage_lora.py. try_load_lora and try_load_lokr
now share this scaffolding alongside the family loaders that follow.
Slices w1a/w2a at the assigned chunk's row range and computes the
partial Hadamard product, mirroring NetworkModuleLokrChunk. Used
when LoHA targets a fused weight (e.g. img_attn.qkv) on models
that expose split Q/K/V modules.
Without self.dim, NetworkModule.calc_scale returns 1.0 even when
alpha is present, silently dropping the trained scaling. Use
w1b.shape[0] as the rank, matching the network_lora and
network_hada conventions.
Vladimir Mandic
Oleksandr Liutyi
CalamitousFelicitousness