NNCF could not compile GPU quantizaiton extensions

[siyhust]

I‘m using nncf 2.6.0 with openvino 2023.3，torch 2.0.1+cu117, ans I installed all packages with pip.

It's ok when I using cpu for PTQ with nncf.quantize，but when I using GPU, it throw a bug below:

"RuntimeError: CUDA is available for pytorch, but NNCF could not compile GPU quantization extensions. Make sure that you have installed CUDA development tools……"

But I indeed install CUDA on /usr/local/cuda-11.7, I also set enviroment with add CUDA path to os.environ['PATH'].

Hope some kind person help me with this question, thanks a lot~

[siyhust]

Is it the reason that I’m using a torch version too low for NNCF ?

[MaximProshin]

@AlexanderDokuchaev , @daniil-lyakhov , please take a look.

[AlexanderDokuchaev]

Hi @siyhust
Version of NNCF an PyTorch are ok, you can find recommended versions here.

Please check that you have installed nvcc compiler by nvcc --version command, may be you have installed only runtime part of cuda.

[siyhust]

@AlexanderDokuchaev
I found it runs ok when I use native terminal instead of Pycharm
Problem fixed by founding correct way to set PATH in my Pycharm.
Thanks a lot.