Install pytorch 0.4.1 with CUDA 10.0 from source

1

Hi, I am trying to install pytorch 0.4.1 with CUDA 10 from source, I follow instructions, but I met some error:

make install -j12
[100%] Generating lib/libnccl.so
make[3]: warning: -jN forced in submake: disabling jobserver mode.
Compiling src/libwrap.cu > /home/yji/pytorch-0.4.1/third_party/build/nccl/obj/libwrap.o
nvcc fatal : Unsupported gpu architecture 'compute_75'
Makefile:137: recipe for target '/home/yji/pytorch-0.4.1/third_party/build/nccl/obj/libwrap.o' failed
make[3]: *** [/home/yji/pytorch-0.4.1/third_party/build/nccl/obj/libwrap.o] Error 1
CMakeFiles/nccl.dir/build.make:60: recipe for target 'lib/libnccl.so' failed
make[2]: *** [lib/libnccl.so] Error 2
CMakeFiles/Makefile2:72: recipe for target 'CMakeFiles/nccl.dir/all' failed
make[1]: *** [CMakeFiles/nccl.dir/all] Error 2
Makefile:129: recipe for target 'all' failed
make: *** [all] Error 2
Failed to run 'bash tools/build_pytorch_libs.sh --use-cuda --use-nnpack nccl caffe2 nanopb libshm gloo THD c10d'

Do you have any idea about this error, I have tried long time to solve it ?
My setup: RTX5000, Ubuntu 18.04, Python 3.6, CUDA 10.0 (I know pytorch 0.4.1 only support CUDA<10.0, but I have to use both CUDA 10.0 and torch 0.4.1)
Thanks in advance.

2

It looks like your cuda version does not support your GPU. You might want to upgrade to latest CUDA. Also make sure that the nvcc that is used is the one with the latest CUDA version.

3

GPU: Quadro RTX 5000, CUDA compiler tool 10.0

CUDA 10.0 should support RTX :sweat_smile:

4

Hmm 7.5 is the latest compute capability, you might want 10.1 or even 10.2.
I couldn’t find the info online. But given the error message, I would bet on this.