Closed akagohary closed 1 month ago
i was wrong on so many levels, i had python 3.12 cuda 12.5 and pytorch 2.4 with no xformers
now to get xformers
you need pytorch 2.3
and cuda 12.1/11.8
, i did the following to make it kinda work on Arch Linux
1 - i fully deleted my venv ( having python 3.12 was a no go and i don't think there is any clean way to downgrade "might be wrong on that ofc since am not python user" )
2- i installed python311
form the AUR
2 - i needed to uninstall my latest cuda,cuda-tools,cudnn
and install cuda 12.1 and its corresponding cuda-tools and cudnn
of the same release age , that requires gcc12
and i had the latest gcc, so you need to install gcc12 from the AUR
3 - after compiling gcc12
"which took about 7 hours on my 6c12t machine" you need to install cuda
manually through downloading them from archlinux archive and using sudo pacman -U cudaXX cudnnXXXX cuda-toolsXXX
4 - now you get to make a new venv in you existing comfyui folder with python version 3.11 , and then i started with pip install xformers==0.0.26.post1
and then installed comfyui requirments folder, that made my comfyui up again
5 - now i had all my plugins broken , but i started to fix them one by one through comfyui manager "running requirments scripts for every addon", but since now i was running python 3.11
it has a cap on what to update and if i run comfyui-3d pack installation again it would just revert what broke it back to earlier version
trying to build wheel for xformers
to my pytorch 2.4
wasn't working on my machine at all and i don't think it generaly works for such a components
now i did try to run comfy-3d-pack using instantmesh and other compination and am running out of cuda memory " i have 12gb vram and 64gb ram " i have no clue actually what is wrong but i though i'd write this before i forget, it has been tiring, i don't think am gonna continue this pursuit cause i'll have to maintain parallel components both on comfyui and my OS itself, untill the whole ecosystem is adapted to a reasonable cuda and pytorch ver i don't think its gonna be a welcoming tinkering environment
PS: you can't just update your system to have the latest stuff cause cuda
will be overitten , there is a way to have parallel cuda's and making a certain version active at a time but i didn't really try it its here >>> https://github.com/mothguib/arch-cuda-parallel
gcc12
and python311
through the AUR @MrForExample thanks for the great work, keep it up =).
hello guys i've tried searching for a solution for my problem but i don't seem to find a fix
running the fix through comfyui manager
my nvcc -V
my pip freeze
this is what i get when installation gets to install.py , both standlone and through comfyui manager
my system is archlinux