Optimizing Quantized Model Inference on ARM with OpenVINO

This project aims to improve INT8 inference performance for quantized YOLO26 models on ARM devices using OpenVINO. Currently, several transformation gaps prevent convolution layers from being executed in INT8, forcing fallback to FP16 kernels and increasing latency. Two key issues were identified. First, patterns such as Conv → Multiply → Add → Swish → FakeQuantize prevent convolution post-op fusion. During the optimization pipeline, the Snippets pass collapses this chain into a subgraph before the graph optimizer can fuse Swish and FakeQuantize into the convolution, preventing INT8 execution. Second, activation zero-points represented as Subtract operations are not correctly folded into quantization parameters, which disrupts precision alignment and contributes to FP16 fallback. This project proposes to address these issues by enabling fusion of Swish and FakeQuantize as sequential convolution post-operations and by refining transformation handling of activation zero-points. Additionally, the optimization pipeline will be adjusted to prevent premature subgraph formation for quantized patterns, allowing graph optimizations to correctly fuse operations. The goal is to maximize INT8 convolution coverage on ARM while preserving numerical correctness.