Hello, amazing work with Guernika, I love using it.

I have a Macbook Air M1 with 8GB of RAM and to run Guernika with XL models I need to compress them. Unfortunately, the conversion fails. After further research it seems that Guernika Model Converter needs to be updated. See this fix Fix quantize-nbits flag.

Currently using version 7.4.1 (1)

Would you provide a new update soon?

Thanks

Diffusers have updated the API, and now it is impossible to convert the SDXL model in the safetensors format.

And whether it is SD1.5 or SDXL, the safetensors format must have a yaml file of the same name next to it, otherwise an error will occur.

huggingface/diffusers#4837

It works normally on MacOS 13.5.

On MacOS 14, The process stayed in the process of saving unet.mlpackage for more than 2 hours, and the ANECompilerService service suddenly appeared, but I chose the GPU.

Running MIL backend_mlprogram pipeline: 100%|██████████| 11/11 [00:00<00:00, 32.32 passes/s]
Saved unet model to /var/folders/rz/mv09yyqj2zs340l11vl_1m080000gn/T/SDXL-Base_unet.mlpackage

I understand that the format used by Apple has some limitations in using LoRas, and right now the best option is to first merge LoRas into a checkpoint and then convert the merged model to mlcore. So a feature for this model converter app to first merge a checkpoint with other checkpoints/LoRas, which is then directly converted would be nice.

It would make my workflow a bit easier. I could use a single app and run it all while I'm doing something else (due to swap file usage, it takes quite long to convert full SDXL checkpoints). I am sure other users are in this situation as well.

In any case, great work on the app! Until earlier today I was doing it with the CLI, this makes things a lot comfier.

hello @GuiyeC :
The initialization of guernikatools packaged with pyinstaller is particularly slow, and you need to wait 1-2 minutes each time before the conversion can start.
I used nuitka to compile the guernikatools package into an executable program. After replacing the GuernikaTools file in GuernikaModelConverter, No need to wait, start converting immediately.
nuitka can convert py into C code for compilation, and it does not release a large number of packaged environments like pyinstaller during runtime, making it faster.
I used the following command to test successfully, but it only works in the local environment:

python3 -m nuitka --assume-yes-for-download --output-dir=./GuernikaModelConverter ./guernikatools/torch2coreml.py

Adding the --standalone parameter can compile the entire running environment to run on other computers, but some errors have been encountered. If you understand, you can research it.