Pytorch 1.5.1 + ubuntu 20.04 + cuda 11

Hello,

I need to test some work on ubuntu 20.04 and cuda 11.
I tried to build pytorch from source and I face just one hiccup easy to fix inside SparseCUDABlas.cu.

But now I am blocking at 89% (so close), and the build failed because of context_gpu.cu.
Does any one have an idea how I can fix this one ?

Thanks


[ 89%] Building NVCC (Device) object caffe2/CMakeFiles/torch_cuda.dir/utils/math/torch_cuda_generated_broadcast.cu.o
[ 89%] Building NVCC (Device) object caffe2/CMakeFiles/torch_cuda.dir/utils/math/torch_cuda_generated_elementwise.cu.o
[ 89%] Building NVCC (Device) object caffe2/CMakeFiles/torch_cuda.dir/utils/math/torch_cuda_generated_reduce.cu.o
/home/sluser/Desktop/bitbucket/hugo/libs/a2kpf/external/torch_cuda/linux/src/c10/util/TypeIndex.h(164): error: expression must have a constant value
/home/sluser/Desktop/bitbucket/hugo/libs/a2kpf/external/torch_cuda/linux/src/c10/util/TypeIndex.h(90): note: expression cannot be interpreted
/home/sluser/Desktop/bitbucket/hugo/libs/a2kpf/external/torch_cuda/linux/src/c10/util/TypeIndex.h(110): note: called from:
detected during:
instantiation of “c10::string_view c10::util::get_fully_qualified_type_name() [with T=caffe2::Tensor]”
/home/sluser/Desktop/bitbucket/hugo/libs/a2kpf/external/torch_cuda/linux/src/c10/util/typeid.h(415): here
instantiation of “c10::string_view caffe2::TypeMeta::TypeName() [with T=caffe2::Tensor]”
/home/sluser/Desktop/bitbucket/hugo/libs/a2kpf/external/torch_cuda/linux/src/caffe2/core/blob.h(88): here

1 error detected in the compilation of “/home/sluser/Desktop/bitbucket/hugo/libs/a2kpf/external/torch_cuda/linux/src/caffe2/core/context_gpu.cu”.
CMake Error at torch_cuda_generated_context_gpu.cu.o.Release.cmake:281 (message):
Error generating file
/home/sluser/Desktop/bitbucket/hugo/build/__torchcuda/caffe2/CMakeFiles/torch_cuda.dir/core/./torch_cuda_generated_context_gpu.cu.o

make[2]: *** [caffe2/CMakeFiles/torch_cuda.dir/build.make:1303: caffe2/CMakeFiles/torch_cuda.dir/core/torch_cuda_generated_context_gpu.cu.o] Error 1
make[2]: *** Waiting for unfinished jobs…
make[1]: *** [CMakeFiles/Makefile2:2807: caffe2/CMakeFiles/torch_cuda.dir/all] Error 2
make: *** [Makefile:161: all] Error 2
CMake Error at tools/cmake/include/HugoPrebuild.cmake:288 (MESSAGE):
[torchcuda] Failed to build
Call Stack (most recent call first):

Reclone the repo recursively again (this sometimes helps).