Release mode cannot find cuda

[dbsxdbsx]

On Win10 with cuda 11.3, and tch-rs 0.7.2 and torchlib 1.11.0_cuda11.3 of release version, when calling tch::Device::cuda_if_available(), in release mode, it turns to using cpu. If it is forced to using cuda by calling tch::Device::Cuda(0), then it runs into error.

[RundownRhino]

@dbsxdbsx Had the same issue. After some searching on the repo, realised this is a duplicate of #291. The solution, as per #291 (comment), is to add

tch::maybe_init_cuda();

before you try to create a Cuda(0) device. That fixes the issue for me.

[dbsxdbsx]

@dbsxdbsx Had the same issue. After some searching on the repo, realised this is a duplicate of #291. The solution, as per #291 (comment), is to add

tch::maybe_init_cuda();

before you try to create a Cuda(0) device. That fixes the issue for me.

@RundownRhino , thanks your advice. After putting tch::maybe_init_cuda(); right before tch::Device::Cuda(0);, now it is not only workable when choosing Cuda, but also workable when choosing cuda_if_available():

  let device = loop {
            let cur_index = get_usize_num_from_user_input_str();
            match cur_index {
                1 => break tch::Device::cuda_if_available(),
                2 => break tch::Device::Cpu,
                3 => {
                    tch::maybe_init_cuda();  // modified here
                    break tch::Device::Cuda(0);
                }
                _ => println!("error input"),
            }
        };

I am wondering why tch::maybe_init_cuda() would affect tch::Device::cuda_if_available().
And still, I think tch::maybe_init_cuda(); is still a workround, there should not need to write this line.

[LaurentMazare]

tch::maybe_init_cuda is certainly a hacky workaround but the best we get so far as afaik it's not easy to specify the linker flags in order and ensure that they will be used to link the final binaries. Closing this issue for now until we find a better way to handle this.