fix: OrtValue.from_dlpack rejects zero-size tensors as non-contiguous

[Copilot]

Description

Fixes IsContiguousTensor in dlpack_converter.cc to correctly accept zero-size tensors, which are vacuously contiguous regardless of their strides.

Root cause: IsContiguousTensor iterates dimensions right-to-left. NumPy 2.x sets all strides to 0 for zero-size tensors. For shape (1, 8, 0, 128), the check hits dim 3 (size=128, stride=0) and fails (128 != 1 && stride[3]=0 != running_size=1) before ever reaching the zero dimension — so the existing in-loop early-return never fires.

Fix: Replace the in-loop zero-dimension check with a pre-scan that short-circuits before any stride validation:

// Zero-size tensors are vacuously contiguous regardless of strides.
// (NumPy 2.x sets all strides to 0 for zero-size tensors.)
for (int i = 0; i < tensor.ndim; i++) {
    if (tensor.shape[i] == 0) {
        return true;
    }
}

Tests: Adds test_ort_value_dlpack_zero_size covering:

• (1, 8, 0, 128) — the KV-cache use case from the issue
• (0,), (0, 4), (4, 0) — other zero-size shapes

Both the arr.__dlpack__() path and the OrtValue round-trip (to_dlpack → from_dlpack) are exercised.

Motivation and Context

OrtValue.from_dlpack raised RuntimeError: ORT only supports contiguous tensor for now for any zero-size NumPy array (e.g. shape (1, 8, 0, 128)), making it unusable for KV-cache models whose initial cache has a zero-length sequence dimension. OrtValue.ortvalue_from_numpy handled the same arrays correctly — only the DLPack path was broken.

Original prompt

---

This section details on the original issue you should resolve

<issue_title>OrtValue.from_dlpack rejects zero-size tensors as non-contiguous</issue_title>
<issue_description>## Describe the issue

OrtValue.from_dlpack rejects zero-size NumPy arrays (e.g. shape (1, 8, 0, 128)) with the error:

RuntimeError: ORT only supports contiguous tensor for now.

Zero-size tensors are vacuously contiguous, so the contiguity check in dlpack_converter.cc is incorrect for this edge case. This matters in practice for KV-cache models where the initial cache has a zero-length sequence dimension.

OrtValue.ortvalue_from_numpy handles zero-size arrays correctly—only the DLPack path is broken.

To reproduce

import numpy as np
import onnxruntime as ort
import onnxruntime.capi._pybind_state as _ort_c

arr = np.zeros((1, 8, 0, 128), dtype=np.float32)

# Works
v1 = ort.OrtValue.ortvalue_from_numpy(arr)
print(v1.shape())  # [1, 8, 0, 128]

# Fails
v2 = ort.OrtValue(_ort_c.OrtValue.from_dlpack(arr.__dlpack__(), False), arr)
# RuntimeError: ...dlpack_converter.cc:242... IsContiguousTensor(dlpack->dl_tensor)
# was false. ORT only supports contiguous tensor for now.

Expected behavior

OrtValue.from_dlpack should accept zero-size tensors, since they are trivially contiguous.

Root cause

IsContiguousTensor in onnxruntime/core/dlpack/dlpack_converter.cc likely fails because the stride computation does not account for dimensions of size 0.

Urgency

Low — easily worked around by falling back to ortvalue_from_numpy for zero-size inputs.

Environment

• onnxruntime: 1.24.2
• Python: 3.13
• OS: Linux (WSL2, x86_64)
  </issue_description>

<agent_instructions>Fix the issue and create good tests</agent_instructions>

Comments on the Issue (you are @copilot in this section)

@justinchuby Wow - this is an issue I asked copilot to create. It does a perfect job. @justinchuby Let me try to also assign copilot

• Fixes OrtValue.from_dlpack rejects zero-size tensors as non-contiguous #27450

---

🔒 GitHub Advanced Security automatically protects Copilot coding agent pull requests. You can protect all pull requests by enabling Advanced Security for your repositories. Learn more about Advanced Security.