ggml-easy
A header-only C++ wrapper around GGML that simplifies loading GGUF or safetensors models, building computation graphs, and running them on CPU or GPU.
ggml-easy is a header-only C++ library that wraps GGML, the tensor library behind projects such as llama.cpp, to reduce the boilerplate needed to load a model and run it. It handles GPU scheduler and buffer setup automatically, with GPU support enabled by default and a fallback to CPU for operations the GPU cannot run.
It can load .safetensors files directly, without first converting them to GGUF, currently for F32, F16 and BF16 tensors, and lets tensors be renamed on load to fit GGML's 64-character name limit. Building a computation graph is done by declaring input and output nodes in a single line each.
It ships with example programs covering basic usage and loading GGUF alongside safetensors for comparison, and it includes helpers for printing intermediate tensor values while debugging a graph.