Leo Fang
The .pth files (`_cuda_bindings_redirector.{pth,py}`) should also be removed.
We probably should pull this in (ex: 12.9.1). Wheel repair (#673) reveals that there are two copies of `msvcp140.dll` getting copied into the wheel ``` ... inflating: cuda/msvcp140-a6d8da6c43532fb4dbf620e6318b0e3c.dll ... inflating:...
> e.g., treating objects with `__cuda_stream__` attribute as `cupy.cuda.Stream`? It would be nice to have, but our preliminary study shows that it is not the most performant thing to do....
> but what would the code be look like when users want to import streams from external library to CuPy? Good question. I would suggest ```python s_torch = torch.cuda.Stream() s_cupy...
ex: https://github.com/cython/cython/blob/8b4c6ac9cd77c98736c687d4db0bc0c84eef624c/Cython/Utility/ObjectHandling.c#L2294-L2297
xref: #9469
@mrogowski could you confirm if LU decomposition has 64-bit index support in cuSOLVER now?