Error when merging SDXL model of any type, including the base model. Tries to make a HUGE file. (SOLVED)

#110
by Cyphex9 - opened

After a fresh install (both A1111 and SDXL), model merging the SDXL base model tries to make a 600gigabyte file!?!?

As the title suggests, I have a fresh install of everything...Git, Python, A1111, cuda, torch, pip, and SDXL, so all are up to date and clean as of 10/25/2023. When I try to merge any SDXL model with any other model, the merge crashes immediately giving me this error code:

"Error merging checkpoints: [enforce fail at ..\c10\core\impl\alloc_cpu.cpp:72] data. DefaultCPUAllocator: not enough memory: you tried to allocate 671088640000 bytes."

Why would it try to make SUCH a huge file?

I have an RTX 4070 gpu (12g), i7-13700k cpu, 32g ram (gskill, 6000MT/s, XMP on), samsung 980 pro SSD, so it feels unlikely (although not impossible) that my system isn't capable. I did see someone mention the 980 might have a caching issue that can be resolved through command arg stuff (which I attempted as well, but without much knowledge on it). I also have very few extensions, and do not have dreambooth installed.

I currently run --autolaunch --no-half-vae --xformers --opt-channelslast in command args. I tried --medvram, and --lowvram in command args just to see...no change. I also tried changing ram to system controlled...no change.

Does anyone know what might cause this? That is a HUGE file it's trying to make.

Happy to offer any additional information that may be needed to diagnose.

Thanks to anyone willing to chime in on this!

(SOLVED?)

I should have figured, but I was told SDXL models run on an entirely different architecture than 1.5, 1.6, 2.0, or 2.1, so model merging between the two isn't even something that is possible to do.

Cyphex9 changed discussion title from Error when merging SDXL model of any type, including the base model. Tries to make a HUGE file. to Error when merging SDXL model of any type, including the base model. Tries to make a HUGE file. (SOLVED)

Sign up or log in to comment