System Info
Please consider reverting #44431
Here is detailed information what problems it causes:
Nerogar/OneTrainer#1506
Besides the one linked, it prevents multiple projects from upgrading or requires workarounds/conversion to match existing LoRA file formats:
| Project |
transformers pin |
Relation to the CLIP flattening (#44431) |
| ai-toolkit |
==5.5.3 |
Pinned one release below the boundary — avoids the flattening entirely. Architecturally closest to OneTrainer (manual loading, kohya-format LoRAs). |
| SimpleTuner |
base deps >=5.6.0 |
Past the boundary, so conversion is necessary — it relies on from_pretrained to convert old↔flat keys and on explicit convert_sd*.py tables for single-file export |
| sd-scripts |
==4.54.1 |
Still on v4 — not affected yet. |
| musubi-tuner |
==4.57.6 |
Still on v4 (same 4.57.6 OneTrainer is on now) — not affected yet. |
Feel free to close this issue if you had very good reasons to do this, but from the outside looking in, it looks like house-keeping that isn't worth it.
Who can help?
text models: @ArthurZucker @Cyrilvallez
System Info
Please consider reverting #44431
Here is detailed information what problems it causes:
Nerogar/OneTrainer#1506
Besides the one linked, it prevents multiple projects from upgrading or requires workarounds/conversion to match existing LoRA file formats:
==5.5.3>=5.6.0from_pretrainedto convert old↔flat keys and on explicitconvert_sd*.pytables for single-file export==4.54.1==4.57.6Feel free to close this issue if you had very good reasons to do this, but from the outside looking in, it looks like house-keeping that isn't worth it.
Who can help?
text models: @ArthurZucker @Cyrilvallez