ggml: Dependency-free C/C++ tensor library with 2-8bit quantization support

A low-level library for handling tensor operations in machine learning. Written in pure C/C++ with no external dependencies, it can be built and run with minimal configuration. It provides predictable performance by avoiding any memory allocation at runtime.

It supports a wide range of architectures, including x86, ARM, RISC-V, LoongArch, PowerPC, s390x, and WebAssembly. In x86, ARM, and RISC-V environments, it improves computational efficiency through SIMD-optimized kernels. It is designed to operate on various backends, including CPUs, GPUs, NPUs, and browsers.

In addition to 2-8bit integer quantization, it supports MXFP4 and NVFP4 microscaling formats. This allows for reduced model size while maintaining inference performance. It is based on the GGUF file format and serves as a core component of the llama.cpp project.

Released under the MIT license, it can be freely applied to both commercial and non-commercial projects. It is suitable for environments that require lightweight tensor operations instead of complex frameworks, particularly in embedded devices or edge computing scenarios. Changes to the core library are recommended to be submitted as pull requests through the llama.cpp repository.