Question about LoRA file identification and usage

[WarriorMama777]

Question about LoRA file identification and usage

Context

I've successfully trained a LoRA model using this repository. However, I'm unsure about which file to use for inference and how to prepare it for use in tools like ComfyUI.

Training Output

After training, I found the following files in the output directory:

• mp_rank_00_model_states.pt (approximately 44GB)
• lora.bin (approximately 20MB)
• Various other files (please let me know if you need information about any other specific files)

Questions

1. Identification of LoRA weights file:

   • Which file contains the actual LoRA weights that should be used for inference?
   • Is it the lora.bin file, or should I be looking at a different file?

2. Usage in inference tools:

   • How can I prepare the LoRA weights file for use in tools like ComfyUI?
   • Is any conversion necessary (e.g., to .safetensors format)?
   • If conversion is needed, could you provide steps or point me to relevant documentation?

3. Clarification on large checkpoint file:

   • What is the purpose of the large mp_rank_00_model_states.pt file?
   • Is this file necessary for using the trained LoRA, or can it be safely ignored for inference?

Thank you in advance for your help in clarifying these points.

[bghira]

when you say success, what do you refer to? it sounds like you've not gotten around to testing it yet?

lora.bin is a pytorch pickle file containing the lora weights. not sure why they didn't use safetensors.

i think you have to send a copy of this file to Comfy for him to add support to CUI.