[BUG]: `__cccl_allocation_stream` invokes cuda-driver UB

[wence-]

Is this a duplicate?

• I confirmed there appear to be no duplicate issues for this bug and that I agree to the Code of Conduct

Type of Bug

Something else

Component

libcu++

Describe the bug

I noticed that cuda::__cccl_allocation_stream holds a static-lifetime cuda::stream

inline ::cuda::stream_ref __cccl_allocation_stream()
{
  static ::cuda::stream __stream{device_ref{0}};
  return __stream;
}

The dtor (and hence cuStreamDestroy) will run below main. But, quoth the documentation

The CUDA interfaces use global state that is initialized during host program initiation and destroyed during host program termination. The CUDA runtime and driver cannot detect if this state is invalid, so using any of these interfaces (implicitly or explicitly) during program initiation or termination after main) will result in undefined behavior.

oops.

How to Reproduce

n/a

Expected behavior

No UB.

Reproduction link

No response

Operating System

No response

nvidia-smi output

No response

NVCC version

No response

[bdice]

One solution is to deliberately allow the resource to leak so that it isn't destroyed after the CUDA context has been shut down, as is done here:

https://github.com/rapidsai/cudf/blob/cb3f0ee50bb21a1821000809edbe59cdd047b0dc/cpp/src/utilities/stream_pool.cpp#L161-L163