Torch::cuda::is_available return false in libtorch 1.5

1

After I update libtorch 1.5(C++, cuda10.1), torch::cuda::is_available() keeps returning false, and I couldn’t find the reason. Before that, I was using libtorch 1.3(C++, cuda10.1), and everything is OK.

2

Which OS are you using?

3

Windows 10 1903, build 18362.778
Nvidia GTX GeForce 1070 GPU

4

You’ll need to pass an additional argument to the linker.
-INCLUDE:?warp_size@cuda@at@@YAHXZ

5

It worked. Thank you very much.

6

I see the same issue with Linux. Is there an equivalent linker magic flag to pass ?

7

Hello,
I have the sample problem with Linux on my NVIDIA Jetson AGX Xavier, JetPack 4.4 and liborch 1.6.
Did you solve this issue on your Linux?
Did you find the “magic flag”?

Thanks