Crash when first convert data type and then flip

[qwerty10086]

🐛 Describe the bug

There is a float tensor in my progarm. When I first convert it to uint16 and then flip, the progarm crashes. I'm not sure whether this is a bug. Here is a minimum example:

#include <torch/torch.h>

int main()
{
    int Size = 4;
    auto T = torch::randn({ Size,Size,1 }, torch::dtype(torch::kFloat).device(torch::kCUDA));

    std::cout << T.flip(0) << std::endl; //No problem
    std::cout << T.to(torch::kInt16).flip(0) << std::endl; //No problem
    std::cout << T.flip(0).to(torch::kUInt16) << std::endl; //No problem
    std::cout << T.to(torch::kUInt16).flip(0) << std::endl; //Crash

    return 0;
}

Versions

Libtorch 2.3.1 with CUDA 11.8

[shink]

FYI, I got "flip_cuda" not implemented for 'UInt16' in this case

[malfet]

Interestingly, it works for CPU, but not CUDA:

>>> import torch
>>> torch.randint(0, 100, (4, 4),  dtype=torch.uint16).flip(0)
tensor([[58, 34, 52,  0],
        [32, 92, 61, 52],
        [53, 33, 17, 64],
        [20, 70, 80, 12]], dtype=torch.uint16)
>>> torch.randint(0, 100, (4, 4),  dtype=torch.uint16, device="cuda").flip(0)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
RuntimeError: "flip_cuda" not implemented for 'UInt16'