Uncompromising Objective

No protobuf

onnx-light replaces protobuf by a custom implementation but keeps the same ONNX format to make it fully compatible. It offers more freedom to implement any custom loading, parsing scenario and speed up this first step (see How-to Python / C++ section).

Compile only what you need

onnx-light is intentionally split into several small C++ libraries (lib_onnx_proto, lib_onnx_op, lib_onnx_lib, lib_onnx_optim, lib_onnx_kernels, lib_onnx_backend_test) so that any downstream project can link only the assembly it actually needs — from a bare proto parser up to the full schema / shape-inference / runtime stack. See How the C++ libraries are split for the detailed breakdown.

C++ Backend Test and Kernels

Backend Tests are implemented in C++. They cannot contain any large tensor and any output is generated through a C++ kernel implemented in C++. The kernels themselves form a self-contained reference implementation of the ONNX operator set, shipped as the lib_onnx_kernels static library: it provides a runtime struct Tensor, RunGraph / RunFunction / RunModel entry points and a per-operator kernel under onnx_light/onnx_kernels/kernels/. The same kernels are used both to compute the expected outputs of every backend test case and to evaluate arbitrary models in C++ without pulling in a third-party runtime (see Using backend tests to evaluate a runtime and Runtime test coverage (onnxruntime and shape_inference)).