i made this script to make my LoRA compatible with nunchaku LoRA loader, as it couldn't handle the text encoder weights from kohya_ss with "train_t5xxl: true".

LoRA Remove Text Encoder Weights

This script filters a LoRA .safetensors file to keep only the U-Net weights, removing any text encoder weights. Useful for sharing or optimizing LoRA files for inference.

Files

• script.py: Main script to filter LoRA weights.
• WINDOWS-install.bat: Sets up a Python virtual environment and installs dependencies (Windows).
• WINDOWS-start.bat: Activates the environment and runs the script (Windows).
• inspect-lora.py: (Optional) Script to inspect LoRA keys.
• lora_keys.txt: (Optional) Example output of LoRA keys.

Quick Start (Windows)

0. open cmd and git clone https://github.com/tazztone/lora-remove-te-weights
1. Place your LoRA file(s) (e.g., m1r4.safetensors) in the PUT-SAFETENSORS-HERE folder.
2. Run WINDOWS-install.bat to set up the environment and install dependencies.
3. Run WINDOWS-start.bat to process your files.

Manual Usage (Other OS / Advanced)

1. Place your LoRA file (e.g., m1r4.safetensors) in this directory or the correct folder.
2. Edit script.py to set the correct input/output filenames or folder if needed.
3. (Recommended) Use a Python virtual environment.
4. Install dependencies:

   pip install torch safetensors numpy

5. Run the script:

   python script.py

Output

• A new file (e.g., m1r4_unet_only.safetensors) containing only U-Net weights.

Notes

• Do not commit large model files (.safetensors) to GitHub.
• Add your virtual environment and model files to .gitignore (see below).

License

MIT