ussoewwin
5526b8cac3
fix(lora): restore v2.2.7 duplicate file I/O elimination
...
Restores the _cached_first_lora_state_dict optimization that was
unintentionally lost during the AWQ modulation layer monkey patch
commit (b618d6f). The cache variable was dropped during the
sectional restructuring of compose_loras_v2().
This eliminates duplicate file I/O for the first LoRA in single-LoRA
workflows, restoring ~40-50% loading time reduction.
2026-05-17 06:53:41 +09:00
ussoewwin
75105eba1c
Key Diffusion implementation: Add key analysis logging to QI LoRA nodes
2026-05-17 05:27:06 +09:00
avan
7375cf98cb
feat(lora_qwen): improve format detection and add safety skip logs
...
- Update _detect_lora_format to support LoRA keys with extra tags like .default.weight.
This ensures compatibility with LoRAs using naming conventions such as:
transformer_blocks.0.attn.add_k_proj.lora_A.default.weight
- Add detailed warning logs to identify which LoRA files had modulation layers (img_mod/txt_mod) skipped for Nunchaku safety.
2026-05-17 05:26:58 +09:00
ussoewwin
3497713d52
Fix: Add Z-Image-Turbo (NextDiT) module resolution fallback paths
...
- Add fallback mappings for NextDiT module naming conventions:
- layers.N.attention.to_qkv -> layers.N.attention.qkv
- layers.N.attention.to_out.0 -> layers.N.attention.out
- layers.N.feed_forward.net.0.proj -> layers.N.feed_forward.w13/w1
- layers.N.feed_forward.net.2 -> layers.N.feed_forward.w2
- Change module resolution failure log level from debug to warning
- Fixes issue where 120 out of 150 LoRA keys were skipped for Z-Image-Turbo models
- Issue was exposed by qwen v2 AWQ fix's stricter module None checks
2026-05-17 05:26:49 +09:00
ussoewwin
dd5969165c
Fix V2 node isolation and zimageturbo_v4 compatibility: Use V2-specific wrapper, add compose_loras_v2_v2 function, fix apply_awq_mod handling, add comfy import
2026-05-17 05:26:49 +09:00
ussoewwin
c837716807
Implement AWQ modulation layer LoRA fix with Runtime Monkey Patch
2026-05-17 05:26:45 +09:00
ussoewwin
c7d8d0c8e0
feat: add AWQ enable switch to V2 node
2026-05-17 05:26:43 +09:00
ussoewwin
614c795f22
Fix Issue #44 : Prevent freeze on unsupported LoRA and optimize loading performance
2026-05-17 05:26:07 +09:00
ussoewwin
d0cb5e3cac
Add optimization reports and update code files
2026-05-17 05:25:12 +09:00
ussoewwin
fb19f541f1
Fix: Improve module resolution for attention and feed_forward layers - Enhanced _get_module_by_name to properly handle hasattr checks and indexable modules - Added NextDiT-specific key mappings for Z-Image-Turbo official loader - Fixed module lookup to correctly find attention.qkv, attention.out, and feed_forward modules - This ensures LoRA weights are applied to all target layers, not just adaLN_modulation
2026-05-17 05:24:57 +09:00
ussoewwin
ad2475c13b
Implement v2.2.3 features: detailed LoRA format detection logs and Qwen V3 toggle UI JS
2026-05-17 05:24:56 +09:00
ussoewwin
f2f98a95e9
Match v2.2.4 AWQ modulation behavior: skip logs, forward-path LoRA for AWQW4A16Linear, and reset support
2026-05-17 05:24:55 +09:00
ussoewwin
c01edb267f
Add v2.2.4 features: AWQ modulation layer detection/skip logic and NextDiT forward signature support
2026-05-17 05:24:55 +09:00
ussoewwin
3bd06a67c3
Restore all deleted files: programs, documentation, and add v2.2.4 changelog
2026-05-17 05:24:50 +09:00
ussoewwin