Why do I need 40GB of RAM to Merge 2 SDXL models?

[okolenmi]

I'm using default simple workflow to merge 2 SDXL models. Why does it consumes so much RAM? Is there a way to do it faster with my 24GB (it's slow when it reaching Swap section)?

[okolenmi]

Claude 3.7 is cooking... I know nothing about safetensors, but AI knows. Claude created an optimized merger for me. I have merged 2 models successfully with 17 GB RAM.

[AzelusLightvale]

If you're willing to go through a bit of a learning curve, comfy-mecha might be up your alley as well. I managed to do a 3 model SDXL merge on 16GB RAM + 4GB VRAM without much issues, and you still get to use Comfy's UI if you want to.
If you can also parse Python code, you could also parse https://github.com/ljleb/sd-mecha/blob/main/sd_mecha/merge_method_wrappers.py for some merge ideas

[ljleb]

Yeah comfy-mecha loads just 1 key at a time from all input checkpoints.

It takes about 7gb (the size of an sdxl checkpoint) to linear interpolate as many models as you like, + the amount of memory needed to hold the largest key in SDXL * the number of input models to ensemble together.