[Afloat (Afloat16)]

Description

Integer compute_data_quant_params uses len(data) to decide whether an array contains values. This raises on a zero-dimensional tensor and misclassifies empty shapes such as (2, 0) as nonempty, leading to invalid reductions or inconsistent override handling.

Use data.size for the three element-count checks. Scalars now follow the same quantization calculation as a one-element vector. Empty integer tensors use the existing identity scale/zero-point defaults regardless of which dimension is empty, while preserving their shape.

Validation

• Seven new unittest methods cover FP16/FP32, INT4/UINT4/INT8/UINT8/INT16/UINT16, symmetric/asymmetric settings, explicit hard-coded scale/code expectations, range overrides, minimum range, readonly inputs and invalid-argument controls.
• A native CPU DequantizeLinear + Add model consumes the scalar quantized initializer after ONNX save/reload.
• New and existing quantization utility tests: 16 methods passed, 279 subtests passed.
• Restoring the original helper produces 1 failing method and 123 failing subtests (pytest reports 124 failures), with 6 completed methods and 25 passing subtests.
• Ruff 0.12.12 lint/format and git diff --check passed.

Exact-source native validation: https://github.com/Afloat16/onnxruntime/actions/runs/43402852485

The runtime was built using the repository's official CPU Release build.py/CMake process at 8dea1194e974fa07b1d2f07a190949c3a92cab80. Its C++ sources match base a8246c0d094f1ab0d6bf236228e5e67683321cd7. This Python-only validation installs that built wheel, overlays the complete quantization package from this exact commit, and verifies byte equality and hashes before normal native imports. It does not claim a separate C++ rebuild for this Python-only change.

Motivation and Context

The public array helpers should determine emptiness by total element count. This change is limited to integer quantization; FLOAT8's separate statistics path is unchanged. Some higher-level initializer callers already flatten their inputs, so this is not a claim that every static quantization entry point previously rejected scalars. No model-quality or throughput improvement is claimed.

[azure-pipelines]

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

[Copilot]

🟢 Approval recommended

No active review issues are reported for the scalar and empty-array quantization changes.

No open issues 🎉

What changed in this PR

This PR fixes scalar and empty-array handling in ONNX Runtime’s integer quantization helpers without changing the FLOAT8 path.

Changes:

• Replace three len(data) checks with data.size, allowing scalars to quantize normally and empty tensors to retain identity parameters and their original shape.
• Add regression tests for quantization types, overrides, readonly inputs, invalid arguments, and scalar serialization followed by native CPU inference.

File	Description
onnxruntime/​test/​python/​quantization/​test_quantize_data_shapes.py	Adds seven test methods covering scalar and empty tensors, controls, and native inference.
onnxruntime/​python/​tools/​quantization/​quant_utils.py	Uses total element count for integer quantization’s emptiness checks.

_{Review effort: Balanced}