Issue with Apex and Pulid : No module named 'fused_layer_norm_cuda'

#6
by Yamiray - opened

Hi, with a fresh install of ComfyUI with UmeAiRT-WAN2.1-AllinOne-Auto_install.bat, I got an issue on workflows containing PulID nodes.
The EVA Loader trigger the following error :
[No module named 'fused_layer_norm_cuda'](ModuleNotFoundError: No module named 'fused_layer_norm_cuda')

I managed to resolve this issue by recompiling Apex by myself using Ninja and the following options : cpp_ext and cuda_ext.

Did other users face the same issue on their own computer ?

same problem

can you explain how_

I used this command (torch 2.7.0 and Cuda 12.8) but it did not work anymore on my computer with the last torch 2.8.0 and Cuda 12.9 :

git clone https://github.com/NVIDIA/apex.git
cd apex
curl -L "https://github.com/NVIDIA/apex/files/14844602/windows_support.patch" | git apply
..\ComfyUI\venv\Scripts\python.exe -m pip install -v --disable-pip-version-check --no-cache-dir --no-build-isolation --config-settings "--build-option=--cpp_ext" --config-settings "--build-option=--cuda_ext" ./

Issue describe here https://github.com/NVIDIA/apex/issues/1792 and here https://github.com/lldacing/ComfyUI_PuLID_Flux_ll/issues/31

I also tried to resolve compilation errrors by creating the following setup.cfg file in the apex directory :
[build_ext]
define = _ENABLE_EXTENDED_ALIGNED_STORAGE

But I got others errors caused by compatibility issues between Nvidia APEX and Cuda 12.9.

From what I read on the Internets, pyTorch include torch.cuda.amp which replace Nvidia Apex with better performances in most cases.
I finaly removed Nvidia APEX to fallback to pyTorch with following command to make PulID work again :
.\ComfyUI\venv\Scripts\python.exe -m pip uninstall apex

Sign up or log in to comment