search

pytorch / vision

Datasets, Transforms and Models specific to Computer Vision
https://pytorch.org/vision
BSD 3-Clause "New" or "Revised" License
15.95k stars 6.91k forks source link

Support for compute capability 8.6 #2958

Closed batrlatom closed 3 years ago

batrlatom commented 3 years ago

Hi, I have tried to compile torchvision for rtx3070 and cuda 11.0, but without any luck. The code I am using is: USE_CUDA=1 USE_CUDNN=1 USE_MKLDNN=1 TORCH_CUDA_ARCH_LIST="8.6" python setup.py install

But I am getting error that this compute capability is not supported. How can I make it compile? Thanks

ptrblck commented 3 years ago

Compute capability 8.6 is supported in CUDA>=11.1, so you would need to update your local CUDA toolkit or compile with 8.0+PTX.

fmassa commented 3 years ago

@ptrblck does it mean that we don't need to do anything in particular in torchvision? Would you say we can close this issue?

ptrblck commented 3 years ago

@fmassa I don't think you need to change anything in torchvision to build with 8.6, as we are doing it in our containers with CUDA11.1 and 8.6 without any changes. I think we can close this issue.

batrlatom commented 3 years ago

ok, thanks for the clarification

vfdev-5 commented 3 years ago

@batrlatom as @ptrblck told me, to compile with 8.0+PTX, you can set TORCH_CUDA_ARCH_LIST="8.0+PTX" with your current cuda 11.0

batrlatom commented 3 years ago

I was able to compile pytorch witch cuda 11.1 and everything works as it should. Thanks!