[TrishaWadhwani]

Description

Fixes #29713.

On the CUDA EP, a device copy scheduled for a weight retained as an in-memory-external OrtValue initializer can have a source and destination that are the same physical host buffer. The host->host branch of GPUDataTransfer::CopyTensor/CopyTensorAsync did not handle this:

• Non-cuda-graph: ORT_ENFORCE(dst_data != src_data) fired -> hard failure.
• CUDA-graph capture: the HOST_ACCESSIBLE cudaStreamSynchronize is illegal during capture -> CUDA 900 -> cudaStreamEndCapture CUDA 901, invalidating the capture.

This became reachable once #29158 (initializer retention) landed.

Fix

Early-return the no-op self-copy at the top of the host->host branch in both methods, before the HOST_ACCESSIBLE stream sync (the key to keeping capture valid). This matches CPUDataTransfer::CopyTensor and the existing GPU->GPU guard.

Testing

Adds a CUDA EP internal regression test (gpu_data_transfer_test.cc) covering sync, async, and cuda-graph-capture self-copies. Verified failing before the fix (ORT_ENFORCE / CUDA 900->901) and passing after, on an RTX 3050 with a CUDA 12.8 sm_86 build.

[azure-pipelines]

Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.

[Tianlei Wu (tianleiwu)]

Nice, focused fix. The key correctness point is right: in CopyTensorAsync the self-copy early-returns before the HOST_ACCESSIBLE cudaStreamSynchronize, which is exactly what keeps a CUDA-graph capture valid (avoids 900 → 901), and it matches CPUDataTransfer::CopyTensor and the existing device→device dst_data != src_data guards. Since the pre-fix path ORT_ENFORCE-crashed on dst == src, there was never any working behavior to preserve, so this strictly removes a failure path with no regression to legitimate copies.

The regression test is thorough — it covers sync, async, and (most importantly) the async self-copy inside a capture region, asserting cudaStreamEndCapture == cudaSuccess.

Only two optional test-hardening nitpicks below; neither blocks. LGTM.

[TrishaWadhwani]

@microsoft-github-policy-service agree

[TrishaWadhwani]

Thanks for the review, Tianlei Wu (@tianleiwu)! Addressed both nitpicks in 5919f6b:

Added explicit assertions that the pinned allocator is HOST_ACCESSIBLE (and the stream's GPU allocator is GPU/DEFAULT) in all three tests, so they fail loudly rather than silently exercising the wrong branch if CreatePreferredAllocators() ordering ever changes.
Reworded the size comment since these tests call CopyTensor{,Async} directly (not through the retention/threshold path), the tensor size doesn't affect which branch runs.
Verified locally: all three tests pass with the fix and fail without it.