ggml src 1d10c0b36ab39a76a1cfa07b008c371f70139a880dd5c478214cbf5556df3cc5 Tensor library for machine learning ggml is a tensor library for machine learning with integer quantization and multi-ISA CPU backends. Optional accelerators are separate packages: * ggml-backend-blas — OpenBLAS * ggml-backend-vulkan — Vulkan * ggml-backend-opencl — OpenCL * ggml-backend-hip — AMD ROCm/HIP (any host CPU with an AMD GPU) Used system-wide by llama-cpp, whisper-cpp, stable-diffusion.cpp and other consumers via find_package(ggml) / *_USE_SYSTEM_GGML. Tensor names are 160 bytes (GGML_MAX_NAME) so Flux/SDXL graphs fit. ComfyUI int8_tensorwise/convrot models use extra I8 kernels on CPU, Vulkan (integerDotProduct) and HIP (hipBLAS I8 GEMM or a portable fallback). Llama/whisper never emit those ops. Runtime library: lib64ggml. Development files: lib64ggml-devel. bero <bero@lindev.ch> https://github.com/ggml-org/ggml