NVCC error compiling fused_bias_act.cpp #72
Comments
Looks like it's picking up CUDA 7.5 version of NVCC. You'd need nvcc to match with whatever CUDA your running with, most likely CUDA 10.0. I'd recommend using the default Docker container template that's provided in the repo. |
|
Thanks, tried with docker and all went smooth :) |
|
This may be a bit ranty, but are there pre-compiled versions? Compilation used to work for me I had to update cuda on my TitanRTX from 10 .1 to 11 and the code no longer works. Driver version 450.119.03 So I made a docker with As awesome are your graphics cards and the idea behind StyleGAN are, the code is a bit weak if it requires local .cu compilation. |
|
@xvdp I understand your ranty position, we're also unhappy that it's difficult to build custom extensions to reliably. I'd recommend completely switching to https://github.com/NVlabs/stylegan2-ada-pytorch which is our official PyTorch StyleGAN2 ADA remake. While it also suffers from tricky configuration issues with building custom extensions, you can use much newer software and containers with it as it works on the latest PyTorch versions. I have done some investigation around making prebuilt custom ops for PyTorch but at least in my experience, there's no clean reliable ABI when building custom ops against PyTorch and CUDA. Thus the prebuilts often break right off the bat if you use them on even a slightly different version of PyTorch than what the extension was built against. I welcome any prior art and examples of successful custom extension binary releases that reliably work across multiple versions of PyTorch and CUDA. |
|
@nurpax Thank you! - I did, I also simplified the projection code so its only one or two calls replacing dlib with mediapipe. Of course that does not answer the compilation issue, but pytorch is more forward compatible than tensorflow. Anycase, it works both on my TitanRTX with Cuda 11 and on the collab. I cant answer as to compilation details, but since StyleGAN has been such an influential project, its functionality deserves to be maintained, cant its sources be maintained in the pytorch repositories? |
I get an error when running
python run_generator.py generate-images ....I solved some of the issues following this https://stackoverflow.com/questions/59342888/tensorflow-error-this-file-requires-compiler-and-library-support-for-the-iso-c#.
Error:
OS info:
GPU info:
nvcc:
g++:
Conda env:
The text was updated successfully, but these errors were encountered: