[BUG]: current_device.create_stream(rmm_stream) segfaults

[bdice]

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

Runtime Error

Component

cuda.core

Describe the bug

I'm seeing a failure with the CUDA stream protocol in cuda-core 0.5.0. This test is now failing where it used to pass.

https://github.com/rapidsai/rmm/blob/00f99c5756ca90b14a48ee203f0e7138193bd722/python/rmm/rmm/tests/test_stream.py#L92-L95

>       buf = cuda_core_mr.allocate(1024, stream=rmm_stream)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
rmm/tests/test_stream.py:91:
------
cuda/core/_memory/_memory_pool.pyx:171: in cuda.core._memory._memory_pool._MemPool.allocate
    ???
------
>   ???
E   TypeError: Stream or GraphBuilder expected, got Stream
cuda/core/_stream.pyx:473: TypeError

Here, I think "got Stream" is really the RMM Stream type which has a stream protocol defined.

@leofang suggested,

We would like users to "launder" their protocol-compliant streams to a cuda.core.Stream like this

dev = Device()
cuda_stream_from_rmm = dev.create_stream(obj=rmm_stream)

and use it in any cuda.core APIs.

I tried that and observed segfaults for both of the following snippets:

rmm_stream = current_device.create_stream(rmm.pylibrmm.stream.Stream())

and

owning_rmm_stream = rmm.pylibrmm.stream.Stream()
rmm_stream = current_device.create_stream(owning_rmm_stream)

How to Reproduce

Run the RMM test linked above.

Expected behavior

No segfault should occur.

Also, it would be nice if a deprecation warning were issued to point users to the create_stream function, with a plan to break the API later on.

Operating System

No response

nvidia-smi output

No response

[bdice]

This was fixed in RMM by adding parentheses:

-        return (0, int(<uintptr_t>self._cuda_stream))
+        return (0, int(<uintptr_t>(self._cuda_stream)))

Thanks @leofang for the catch! rapidsai/rmm#2187 should help eliminate this class of potential issues with Cython casts for the future.