cannot initialize CUDA without ATen_cuda library

[pharrellyhy]

Just want to test if CUDA part is working properly using below code but got errors. I linked both cpu and cuda libraries. The cpu part is working fine. Any idea on this?

cout << ones(CUDA(kFloat), {3,4}) << "\n";

TARGET_LINK_LIBRARIES(test_aten ${CUDA_LIBRARIES} ${ATen_BINARY_DIR}/src/ATen/libATen_cuda.so ${ATen_BINARY_DIR}/src/ATen/libATen_cpu.so)

terminate called after throwing an instance of 'at::Error'
what(): cannot initialize CUDA without ATen_cuda library (initCUDA at /home/pharrell/downloads/ATen/aten/src/ATen/detail/CUDAHooksInterface.h:42)
frame #0: at::Context::lazyInitCUDA()::{lambda()#1}::operator()() const + 0x32 (0x409a56 in ./test_aten)
frame #1: void std::_Bind_simple<at::Context::lazyInitCUDA()::{lambda()#1} ()>::_M_invoke<>(std::_Index_tuple<>) + 0x28 (0x40afb0 in ./test_aten)
frame #2: std::_Bind_simple<at::Context::lazyInitCUDA()::{lambda()#1} ()>::operator()() + 0x2c (0x40ace0 in ./test_aten)
frame #3: void std::__once_call_impl<std::_Bind_simple<at::Context::lazyInitCUDA()::{lambda()#1} ()> >() + 0x17 (0x40a7a4 in ./test_aten)
frame #4: + 0xea99 (0x7fe48d128a99 in /lib/x86_64-linux-gnu/libpthread.so.0)
frame #5: ./test_aten() [0x408ccd]
frame #6: void std::call_onceat::Context::lazyInitCUDA()::{lambda()#1}(std::once_flag&, at::Context::lazyInitCUDA()::{lambda()#1}&&) + 0x77 (0x40a0f7 in ./test_aten)
frame #7: at::Context::lazyInitCUDA() + 0x3d (0x409b37 in ./test_aten)
frame #8: at::Context::initCUDAIfNeeded(at::Backend) + 0x21 (0x409b81 in ./test_aten)
frame #9: at::Context::getTypeOpt(at::Backend, at::ScalarType) + 0x23 (0x40984b in ./test_aten)
frame #10: at::Context::getType(at::Backend, at::ScalarType) + 0x4a (0x40992a in ./test_aten)
frame #11: ./test_aten() [0x408da9]
frame #12: ./test_aten() [0x408de1]
frame #13: main + 0x65a (0x409467 in ./test_aten)
frame #14: __libc_start_main + 0xf0 (0x7fe48c3cc830 in /lib/x86_64-linux-gnu/libc.so.6)
frame #15: _start + 0x29 (0x408bb9 in ./test_aten)

Aborted (core dumped)

[c-hofer]

I had a similar problem which was solved by loading the cuda verson of ATen library in my executable:

Add a line in ure main like the following

#include <dlfcn.h> //necessary to use dlopen
...
int main(){
dlopen("libcaffe2_gpu.so", RTLD_NOW);
...
}

and ensure that LD_LIBRARY_PATH makes it visible on runtime.

Note: maybe the library has a different name in pytorch 0.5.0 this one seems to be right (at least at the moment)
regards

[pharrellyhy]

@c-hofer
Thanks. This works like a charm. Closing it now.