Integer-kernel deployment of QuAR

Develop an integer-kernel deployment of Quantizer-Aligned Recalibration (QuAR) and characterize its latency, memory, and energy on an integer-only inference stack.

Background

QuAR is evaluated as a single-pass, parameter-free adaptation method for post-training-quantized vision transformers, but the reported implementation uses an explicit per-channel multiply/add wrapper around frozen quantizers. Although the paper verifies that this wrapper is representable in ONNX, it does not implement the method within an end-to-end integer-only inference kernel. The unresolved deployment problem is therefore to determine the practical latency, memory, and energy costs of QuAR on such hardware and software stacks.