The llama.cpp project has added support for the Tencent Hunyuan 3 (HYV3ForCausalLM) model architecture, designated as hy_v3 in GGUF format. This update enables users to run this specific MoE decoder stack within the library.
- Adds full hy_v3 architecture support, including per-head Q/K RMSNorm and a sigmoid router with expert selection bias.
- Implements an always-active ungated shared expert and leading dense blocks via first_k_dense_replace.
- Ported from charlie12345's fork and adapted to current mainline APIs for compatibility with existing hy_v3 GGUFs.
This change allows the community to utilize Tencent Hunyuan 3 models on various platforms, including macOS, Linux, Windows, and Android, across diverse hardware backends like CUDA, ROCm, and Vulkan.