Build pytorch from source with cuda

deployment
1

I built pytorch from source, infos:
* centos7
* pytorch v1.8.1
* cuda 10.2
* cudnn 8.4.0
But after installation, libtorch_cuda.so is quite different from the official whl. The version I compiled need “libcusparse.so.10, libcurand.so.10, libcusolver.so.10, libcudnn.so.8, libcufft.so.10, libcublas.so.10”. How can I get a libtorch_cuda.so like the official version?

image
image713×351 8.07 KB

image
image701×260 7.05 KB

2

I resolved it by set option(CAFFE2_STATIC_LINK_CUDA “Statically link CUDA libraries” ON). In addition, cudnn version should < 8.3.0 when compiling statically.