Fix float16 QuantizeLinear rounding on CPU

[kadyrbekovhamit-cyber]

Description

Fixes the still-reproducible numerical issue reported in #18576.

The CPU MLFloat16 overload of ParQuantizeLinearStd converted x / scale directly with static_cast<int32_t>, which truncates toward zero. ONNX QuantizeLinear requires round-to-nearest with ties to even.

This change applies std::nearbyint before the integer conversion, matching the ONNX contract and the nearby blocked float16 paths in the same file.

Minimal reproduction

For opset 19:

x (float16)     = [ 0.050018310546875, -0.050018310546875]
scale (float16) = 0.0999755859375
zero_point      = 0 (int8)
expected        = [ 1, -1]
ORT CPU 1.29.0  = [ 0,  0]

NumPy nearest-even and ONNX ReferenceEvaluator both return [1, -1]. A 4,096-value float16 differential check found 2,041 mismatches, all toward zero and by at most one quantization unit. The float32 control path matched the oracle.

Changes

• round the scalar float16 CPU path with std::nearbyint;
• add a non-integral float16-to-int8 regression test.

Testing

• reproduced on macOS ARM64 with onnxruntime==1.29.0 and onnx==1.22.0;
• verified against NumPy and ONNX ReferenceEvaluator;
• full C++ unit suite not run locally; the new OpTester case is included for CI.

This is a numerical-correctness fix, not a security change.

[azure-pipelines]

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

[Copilot]

🟡 Changes recommended

The test does not yet verify exact ties-to-even behavior.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

Fixes CPU float16 QuantizeLinear to round fractional values before integer conversion.

Changes:

• Applies nearest rounding in the scalar float16 quantization path.
• Adds a regression test for positive and negative fractional values.

File summaries

File	Description
onnxruntime/core/util/qmath.h	Corrects float16 quantization rounding.
onnxruntime/test/providers/cpu/tensor/quantize_linear_test.cc	Adds regression coverage.

Review details

• Files reviewed: 2/2 changed files
• Comments generated: 1
• Review effort level: Balanced

---

💡 Configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[kadyrbekovhamit-cyber]

@microsoft-github-policy-service agree

[Copilot]

🟢 Approval recommended

The focused implementation matches existing quantization paths and is covered by an appropriate regression test.

Review details

• Files reviewed: 2/2 changed files
• Comments generated: 0 new
• Review effort level: Balanced