anyone got this working on windows
if working on windows - what is your system i.e. python , pytorch and cuda version
I have pytorch 2.11, cuda 13, and python 3.11 in portable. I do have python 3.12 in my system so idk if that plays a part but its there.
I cannot run THIS version of the model but i merged the fp8 shards into a safetensor file and curated the custom nodes to fit my model.
Again, cant test because of OOM but im pretty sure it runs. Your welcome to test it, it fails after first sampling steps for me. Could be the nodes, the bridge from model to sr model maybe be broken but im not sure. I didnt have alot of time to troubleshoot last night but i got pretty far. If you test it, let me know your results!
https://huggingface.co/realrebelai/DaVinci_MagiHuman_fp8_merges
Be sure to use MY fork for the custom nodes linked in description, smthemex's nodea only house the bf16, mine incorporate the fp8 architecture.
Be sure to use MY fork for the custom nodes linked in description, smthemex's nodea only house the bf16, mine incorporate the fp8 architecture.
but i cant even get the magihuman node working - tried many versions of python , pytorch and cuda
Yea its everybody right now, theres alot more to this model than original anticipated. The architecture behind it creates alot of issues when you change it. I thino it has something to do with how it uses the magicompiler but im not sure. Ive been editing the nodes for 3 days now and i still cant get a generation out either.
I think i cracked the code tbh, you can test my setup, need 30gb combined compute
https://github.com/RealRebelAI/ComfyUI_MagiHuman_fp8_ditFIX_nodes