[mirounga]

…uantizeLinear

A16 models keep 16-bit activation Q/DQ boundaries around every quantized op, including the MatMulNBits nodes fused in #32885. On CPU each pair runs as two kernels and materializes an int16/uint16 tensor that is immediately dequantized again.

Add the com.microsoft.QuantizeDequantizeLinear op (float input, int16 or uint16 zero point) and a CPU kernel for it:

• MlasQuantizeDequantizeLinear runs the per-element steps of the MLAS QuantizeLinear kernel and then (q - zp) * scale in one vector loop (SSE2, NEON, LSX). On other platforms, or when a non-default x64 quantize kernel is selected, it falls back to two passes through MlasQuantizeLinear.
• Results are bitwise identical to the unfused QuantizeLinear -> DequantizeLinear pair.

Add QuantizeDequantizeLinearFusion, a Level3 CPU transformer that replaces a 16-bit QuantizeLinear whose consumers are all DequantizeLinear nodes with the same constant scalar scale and zero point by one fused node. It also follows the quantized tensor through data-movement ops (Transpose, Reshape, Slice, Gather, Split, Unsqueeze, ...) to matching DQs and runs those ops on the float values. It runs after the Level2 Q/DQ node-unit fusions, so it never removes a Q/DQ they could still use, and session.disable_quant_qdq disables it.

Tests:

• QuantizeDequantizeLinearOpTest: int16 and uint16 round-half-to-even and saturation, a large tensor against a reference, empty input, and rejection of per-axis parameters.
• MLAS QuantizeDequantizeLinearS16/U16: bitwise comparison with MlasQuantizeLinear followed by dequantization.
• QuantizeDequantizeLinearFusionTests: interior pairs are bit-exact, fusion through data-movement ops, fan-out and graph-output DQs, A16 MatMulNBits boundaries, session.disable_quant_qdq, and the cases that must not fuse (8-bit, mismatched or non-constant or per-axis parameters, other uses of the quantized value, non-CPU EPs).

[Copilot]

🟡 Changes recommended

The fusion accepts opset-23+ QuantizeLinear nodes with FLOAT16 precision but computes in FLOAT, changing model outputs.

1 open issue

• Reject FLOAT16 precision for fused QuantizeLinear · New

What changed in this PR

This PR adds CPU fusion for 16-bit QuantizeLinear → DequantizeLinear pairs, reducing intermediate tensor storage and kernel overhead in A16 models, including the MatMulNBits boundaries preserved by #32885.

Changes:

• Adds the com.microsoft.QuantizeDequantizeLinear operator, CPU kernels, and MLAS implementation.
• Registers a Level3 fusion that follows compatible data-movement operations and respects session.disable_quant_qdq.
• Adds numerical and graph-transformation tests, plus operator documentation.

File	Description
onnxruntime/​test/​optimizer/​quantize_dequantize_linear_fusion_test.cc	Tests fusion patterns, exclusions, and A16 boundaries.
onnxruntime/​test/​mlas/​unittest/​test_quantizelinear.cpp	Adds bitwise comparisons against separate quantization and dequantization.
onnxruntime/​test/​contrib_ops/​quantize_dequantize_linear_op_test.cc	Tests rounding, saturation, empty inputs, and parameter validation.
onnxruntime/​core/​optimizer/​qdq_transformer/​quantize_dequantize_linear_fusion.h	Declares the fusion transformer.
onnxruntime/​core/​optimizer/​qdq_transformer/​quantize_dequantize_linear_fusion.cc	Implements eligibility checks and graph rewriting.
onnxruntime/​core/​optimizer/​graph_transformer_utils.cc	Registers the fusion at Level3.
onnxruntime/​core/​mlas/​lib/​quantize.cpp	Implements vectorized fusion and two-pass fallback.
onnxruntime/​core/​mlas/​inc/​mlas.h	Declares the fused MLAS API.
onnxruntime/​core/​graph/​contrib_ops/​quantization_defs.cc	Defines the operator schema and shape inference.
onnxruntime/​core/​graph/​contrib_ops/​ms_opset.h	Registers the schema in the Microsoft operator set.
onnxruntime/​contrib_ops/​cpu/​quantization/​quantize_dequantize_linear.cc	Implements the threaded CPU operator.
onnxruntime/​contrib_ops/​cpu/​cpu_contrib_kernels.cc	Registers int16 and uint16 CPU kernels.
docs/​OperatorKernels.md	Lists supported CPU kernel types.
docs/​ContribOperators.md	Documents the new operator.

_{Review effort: Balanced}