[BUG]: PinnedMemoryResource returns an unusable buffer when host memory pools are unsupported

[uday1o1]

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

On a device where device.properties.host_memory_pools_supported is False, PinnedMemoryResource.allocate() returns a buffer instead of reporting that host memory pools are unsupported.

The first host-to-device transfer using the returned buffer then fails with CUDA_ERROR_INVALID_VALUE.

Synchronizing the stream immediately after allocation does not expose the failure earlier or change the result.

LegacyPinnedMemoryResource completes the same allocation and transfer sequence successfully.

This makes the transfer appear invalid even though the underlying limitation is that the stream's device does not support host memory pools.

How to Reproduce

1. Use a CUDA device where host_memory_pools_supported is False.
2. Allocate pinned memory with PinnedMemoryResource.
3. Copy from the returned buffer to device memory.
4. Observe CUDA_ERROR_INVALID_VALUE.

Expected behavior

PinnedMemoryResource.allocate() should report that the stream's device does not support host memory pools instead of returning an unusable buffer.

The documentation should state this requirement and recommend LegacyPinnedMemoryResource when host memory pools are unavailable and stream-ordered allocation is not required:

• PinnedMemoryResource documentation
• LegacyPinnedMemoryResource documentation

System information

Operating System

• Jetson AGX Orin
• Jetson Linux R39.2
• Python 3.12.3
• CUDA 13.2.1
• cuda-core 1.1.0
• cuda-bindings 13.3.1