[Justin Chu (justinchuby)]

Description

Modernize public/native OrtValue DLPack interchange, remove obsolete export helpers, and retain legacy capsule imports.

Fix NumPy ownership, CUDA/ROCm handoffs, and exception-safe batched exports. Add mandatory inference-CI coverage for dtype compatibility and lifetime/error paths.

Motivation and Context

Follow-up to #33848. Make tensor-object interchange safe for framework use, including PyTorch's packed-type limits.

[Copilot]

🟢 Approval recommended

No active review issues are recorded for the inference-only PyTorch DLPack interoperability changes.

No open issues 🎉

What changed in this PR

Adds inference-only PyTorch DLPack interoperability coverage following #33848, validating the existing dtype support without requiring ORT training.

Changes:

• Tests bidirectional exchange for 18 byte-aligned dtypes, all float8 bit patterns, and compatibility boundaries for five packed dtypes, including zero-copy sharing and ownership.
• Runs the suite in transformers-enabled Python test jobs when contrib operators are enabled.
• Removes a redundant OrtValue copy, simplifies shape access, and clarifies ownership and PyTorch compatibility documentation.

File	Description
tools/​ci_build/​build.py	Adds mandatory DLPack suite execution to eligible test jobs.
onnxruntime/​test/​python/​onnxruntime_test_python_dlpack.py	Adds dtype, storage, lifetime, and layout interoperability tests.
onnxruntime/​python/​onnxruntime_pybind_state_common.h	Declares the export helper with a reference parameter and clarifies ownership.
onnxruntime/​python/​onnxruntime_pybind_state_common.cc	Removes the export helper’s redundant OrtValue copy.
onnxruntime/​python/​onnxruntime_pybind_ortvalue.cc	Clarifies capsule ownership in binding documentation.
onnxruntime/​python/​onnxruntime_inference_collection.py	Documents buffer lifetime and PyTorch dtype compatibility.
onnxruntime/​core/​dlpack/​dlpack_converter.h	Corrects the converter’s ownership comment.
onnxruntime/​core/​dlpack/​dlpack_converter.cc	Reuses the dimension span for shape access and stride reservation.

_{Review effort: Balanced}

[Copilot]

🔵 Needs a closer look

An existing dlpack property raising AttributeError is still masked by a generic TypeError; descriptor failures must be preserved and covered by regression tests for both importers.

No open issues 🎉

1 previously missed issue

In code that hasn't changed since last review

Preserve AttributeError raised by existing __dlpack__ descriptors

onnxruntime/​python/​onnxruntime_pybind_state_common.cc:70

If an existing __dlpack__ property raises AttributeError, this branch clears the producer's exception and replaces it with the generic invalid-capsule TypeError. That still suppresses lookup failures despite the stated error-preservation contract. Distinguish an absent protocol attribute from a failing descriptor, preserve the latter exception, and add an AttributeError-raising property regression for the public and native importers.

_{Review effort: Balanced}

[Copilot]

🟡 Changes recommended

The new CUDA provider method uses CUDA_RETURN_IF_ERROR before it is defined, causing CUDA builds to fail compilation.

3 open issues

• Use defined CUDA error checks in the new method · New
• Include MIGraphX interface builds in provider lookup guard · New
• Request the no-sync DLPack stream for GPU OrtValues · New

_{Review effort: Balanced}

[Copilot]

🟡 Changes recommended

The DLPack bindings hold the GIL during device-wide synchronization, blocking other Python threads, and need to release it safely with exclusive capsule ownership established first for imports.

1 open issue

• Release the GIL during DLPack device synchronization · New

3 resolved since last review

• Use defined CUDA error checks in the new method
• Request the no-sync DLPack stream for GPU OrtValues
• Include MIGraphX interface builds in provider lookup guard

_{Review effort: Balanced}

[Copilot]

🟡 Changes recommended

The DLPack import fails in PyTorch 2.10 when the producer tensor’s CUDA device differs from the current device, also causing the new protocol-path test to fail on multi-GPU systems.

1 open issue

• DLPack import fails when the producer CUDA device isn't current · New

1 resolved since last review

• Release the GIL during DLPack device synchronization

_{Review effort: Balanced}

[Copilot]

🟡 Changes recommended

The change fails to compile with Abseil disabled, risks exhausting memory by retaining NumPy allocations while the main thread is blocked, and lacks successful ROCm/MIGraphX handoff coverage.

3 open issues

• Avoid pair keys unsupported by std::hash in the no-Abseil build · New
• Queue NumPy releases on the cleanup thread when the main thread is blocked · New
• Add ROCm coverage for GPU handoff ordering and device restoration · New

1 resolved since last review

• DLPack import fails when the producer CUDA device isn't current

_{Review effort: Balanced}

[Copilot]

🔵 Needs a closer look

Native ownership cleanup and GPU stream/device restoration require runtime validation on matching CUDA and ROCm builds, including multi-GPU and graph-capture paths, which this review environment cannot provide.

No open issues 🎉

3 resolved since last review

• Avoid pair keys unsupported by std::hash in the no-Abseil build
• Add ROCm coverage for GPU handoff ordering and device restoration
• Queue NumPy releases on the cleanup thread when the main thread is blocked

_{Review effort: Balanced}

[Copilot]

🔵 Needs a closer look

Synchronization and ownership changes remain unverified because matching CUDA/ROCm/MIGraphX hardware was unavailable to validate multi-device restoration, graph-capture safety, and cross-thread cleanup.

No open issues 🎉

_{Review effort: Balanced}