Summary

• Add __dlpack__, __dlpack_device__, from_dlpack(), and __array__() to the Python OrtValue wrapper class
• Enable zero-copy tensor sharing between ONNX Runtime and DLPack-compatible frameworks (PyTorch, JAX, CuPy)
• Allow np.array(ort_value) and np.asarray(ort_value) via the numpy array protocol

Motivation

Fixes #24071

The C++ pybind layer (C.OrtValue) already implements __dlpack__, __dlpack_device__, and from_dlpack, but these aren't exposed on the user-facing onnxruntime.OrtValue class. Users currently have to reach through ort_value._ortvalue to access DLPack functionality. Similarly, there's no __array__ protocol, so np.asarray(ort_value) doesn't work.

This change surfaces these protocols on the public API so that:

# DLPack: zero-copy sharing with PyTorch
import torch
torch_tensor = torch.from_dlpack(ort_value)
ort_value = OrtValue.from_dlpack(torch_tensor)

# numpy: transparent conversion
arr = np.asarray(ort_value)

Changes

• onnxruntime/python/onnxruntime_inference_collection.py: Added four methods to OrtValue:

  • __dlpack__(stream=None) — delegates to C.OrtValue.__dlpack__
  • __dlpack_device__() — delegates to C.OrtValue.__dlpack_device__
  • from_dlpack(source) — classmethod that calls source.__dlpack__(), auto-detects bool tensors via is_dlpack_uint8_tensor, and creates an OrtValue
  • __array__(dtype=None, copy=None) — delegates to numpy() with optional dtype casting
  • All DLPack methods guard against builds where DLPack is disabled

• onnxruntime/test/python/onnxruntime_test_python.py: Added 5 tests:

  • test_ort_value_wrapper_dlpack — round-trip via __dlpack__/__dlpack_device__/from_dlpack
  • test_ort_value_from_dlpack_numpy — OrtValue.from_dlpack(numpy_arr)
  • test_ort_value_from_dlpack_dtypes — dtype preservation across float32/float64/int32/int64
  • test_ort_value_array_protocol — np.array()/np.asarray() with dtype casting
  • test_ort_value_array_protocol_int — integer tensor via __array__

Test Plan

• All 5 new tests pass locally
• Existing test_ort_value_dlpack and test_ort_value_dlpack_zero_size tests remain unaffected
• No C++ changes; pure Python addition to the wrapper layer
• ruff check and ruff format pass clean