Agent: Custom C++, AMD ROCmLLM: Llama, Qwen#quantization#amd#rocm#llama.cpp#inference-optimization
ROCmFPX extends llama.cpp with AMD-optimized 2-8 bit quantization formats, featuring HIP/ROCm and Vulkan acceleration. Achieves 18.5% speed improvements on Strix Halo with ROCmFP2 format while maintaining model quality through advanced codebook techniques.