[adrastogi]

Description

Fixes the two async same-buffer tests added in #29792 (onnxruntime/test/providers/cuda/test_cases/gpu_data_transfer_test.cc):

• GPUDataTransferSameBufferTest.CopyTensorAsyncSelfCopyIsNoOp
• GPUDataTransferSameBufferTest.CopyTensorAsyncSelfCopyDuringCaptureKeepsCaptureValid

Both constructed CudaStream with own_flag=false and nullptr external cuBLAS/cuDNN handles. In that mode CudaStream::CudaStream (core/providers/cuda/cuda_stream_handle.cc) calls cublasSetStream(cublas_handle_, stream) on the null handle, which throws:

CUBLAS failure 1: the library was not initialized ... expr=cublasSetStream(cublas_handle_, stream)

Test-only change:

• Construct both streams with own_flag=true (matching cuda_execution_provider_test.cc) so CudaStream creates its own cuBLAS (and cuDNN, if enabled) handle.
• Remove the explicit cudaStreamDestroy(cuda_stream) calls, because an owning CudaStream destroys the stream in its destructor (avoids a double destroy).
• The capture test still keeps the stream alive through cudaStreamEndCapture and graph destruction. The first test still synchronizes the stream before scope exit.

No product code changes. GPUDataTransfer::CopyTensorAsync does not call CudaStream::Flush(), so owning the stream doesn't add a capture-illegal synchronize.

Motivation and Context

Fixes #33930. These tests fail on main in Test Linux CUDA x64 Release, Test Linux TensorRT x64 Release, and Windows GPU CUDA CI Pipeline Test Job.

Validation

• clang-format --dry-run --Werror onnxruntime/test/providers/cuda/test_cases/gpu_data_transfer_test.cc (VS clang-format 22.1.3): passed.
• I didn't run the CUDA tests locally because this machine has no NVIDIA GPU. Validation depends on the CUDA CI jobs above: Test Linux CUDA x64 Release, Test Linux TensorRT x64 Release, and Windows GPU CUDA CI Pipeline Test Job.

[Copilot]

🟡 Changes recommended

Removing explicit stream cleanup leaks CUDA streams in both async tests under USE_CUDA_MINIMAL because CudaStream’s destructor omits cudaStreamDestroy in that configuration.

1 open issue

• Restore explicit stream cleanup in CUDA-minimal async tests · New

What changed in this PR

This test-only PR addresses #33930 by fixing stream ownership in the two asynchronous same-buffer copy regression tests added in #29792. No product code changes are included.

Changes:

• Set own_flag=true so each CudaStream creates valid library handles.
• Remove explicit stream destruction in favor of destructor cleanup.
• Preserve synchronization and capture-test stream lifetimes.

File	Description
onnxruntime/​test/​providers/​cuda/​test_cases/​gpu_data_transfer_test.cc	Updates stream ownership and cleanup in both async same-buffer copy tests.

_{Review effort: Balanced}

[Copilot]

🔵 Needs a closer look

GPU execution was not validated locally, so both async tests, including the capture-validity assertion, still need confirmation in Linux CUDA, TensorRT, and Windows CUDA CI.

No open issues 🎉

1 resolved since last review

• Restore explicit stream cleanup in CUDA-minimal async tests

_{Review effort: Balanced}

[adrastogi]

[Agentic Response]: Closing as superseded. #33916 already landed the same test fix on main (own_flag=true and no manual cudaStreamDestroy in both async GPUDataTransferSameBufferTest tests). This PR's only extra change was an explicit stream destroy under USE_CUDA_MINIMAL, which only matters for CUDA-minimal builds with the CUDA EP internal tests enabled. Not worth a separate PR for now.