-
Notifications
You must be signed in to change notification settings - Fork 12
Compiling modules on Linux #12
Comments
I get this issue when I'm trying to use the latest branch with SDXL. Any ideas? |
@CyberTimon |
Thank you this fixed this issue. Now hopefully the last issue:
Do you know how I should fix this? And the filename of the downloaded unet is still very random |
After furter investigation I noticed that when I first load up comfy with the node it downloads the aitemplate unet and it outputs the first issue (file too short), even after changing blob to resolve. The second time I execute the prompt I get the above input format error. |
Ah, yes the the script intends them to be compressed for the download. The compressed versions will be uploaded later. Instead, download the module yourself, here is the direct link and rename it to the same as the existing file |
Ahhh it really seems impossible to use AiTemplate on linux on my system. I did exactly what you said. I now have a single
|
See #15 |
Thank you. I renamed the suggested things and now I get this:
I think it's the best to wait until you know a real fix or a PR appears. |
Same as @CyberTimon, here's the full error:
|
This issue can serve as a discussion point for any issues related to compiling modules on Linux
My setup is as follows:
I use WSL, with Ubuntu 20.04
CUDA installation is specific to WSL, for other distros see cuda-downloads.
Earliest known working CUDA toolkit version is 11.6, I'd recommend using the latest
Ubuntu 20.04 includes Python 3.8.10, minimum Python version is 3.8, and I use PyTorch nightly CUDA 12.1, PyTorch version working at time of AITemplate release was 1.12.1, so anything newer than that should work
The following instructions use the build of AITemplate mentioned here, it includes the pending PRs, it is set up for CMake so that is disabled here, CUDA path is set so nvcc is found, a venv is used
To note, using that build is not necessary, and it will eventually become outdated, Installation from source is like
Until the pending PRs mentioned are merged they will need to be merged to a local branch before installation, or the changes applied to the installed package.
Here's some SDXL UNet modules I compiled earlier, covering 1024, 1536, 2048 max resolution, batch size 1, sm80
unet_xl_linux.zip
This is actually a
7z
archive rather than.zip
due to GitHub's file extension constraints.cc @FizzleDorf These UNet modules are ready uploaded to Huggingface and added to modules.json, existing Linux VAE modules will work for SDXL with the fp16 fixed version
The text was updated successfully, but these errors were encountered: