Skip to content

Apply LoRA during model conversion #709

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom
Draft
wants to merge 1 commit into
base: master
Choose a base branch
from
Draft

Apply LoRA during model conversion #709

wants to merge 1 commit into from
+78 −16

Conversation

Green-Sky
Copy link
Contributor

@Green-Sky Green-Sky commented Jun 29, 2025 •
edited
Loading

Very rough code, but works.
Useful for avoiding the lora overhead during inference, or just mixing your own model from loras.

Patch based on (outdated) @stduhpf / bleedingedge : Green-Sky@69028c7

Keep in mind that the weights are first quantized, and then it tries to apply the loras. (which is very not ideal)

Example

merging lcm lora into a sd1.5 finetune:

$ result/bin/sd -M convert -m models/CyberRealistic_V9_FP16.safetensors --lora-model-dir models/loras/sd1/ -p "<lora:lcm-lora-sdv1-5:1>" -o models/CyberRealistic_V9-lcm-q8_0.gguf
[INFO ] model.cpp:910  - load models/CyberRealistic_V9_FP16.safetensors using safetensors format
[INFO ] model.cpp:2106 - lora lcm-lora-sdv1-5:1.00
[INFO ] model.cpp:2121 - found at 'models/loras/sd1/lcm-lora-sdv1-5.safetensors'
[INFO ] model.cpp:1987 - model tensors mem size: 2035.18MB
  |==================================================| 1131/1131 - 500.00it/s
[INFO ] model.cpp:2032 - load tensors done
[INFO ] model.cpp:910  - load models/loras/sd1/lcm-lora-sdv1-5.safetensors using safetensors format
[INFO ] lora.hpp:117  - loading LoRA from 'models/loras/sd1/lcm-lora-sdv1-5.safetensors'
  |==================================================| 834/834 - 0.00it/s
[INFO ] model.cpp:2047 - applied 'models/loras/sd1/lcm-lora-sdv1-5.safetensors':1.000000
[INFO ] model.cpp:2052 - trying to save tensors to models/CyberRealistic_V9-lcm-q8_0.gguf
convert 'models/CyberRealistic_V9_FP16.safetensors'/'' to 'models/CyberRealistic_V9-lcm-q8_0.gguf' success

@Green-Sky
Copy link
Contributor Author

ggml switched to c++17, sd.cpp should too.

@Green-Sky Green-Sky changed the title merge in loras during model conversion Apply LoRA during model conversion Jun 29, 2025
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers
No reviews
Assignees
No one assigned
Labels
None yet
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
1 participant
@Green-Sky