llama.cpp

feat: MTP support for dense Qwen 3.5 with FastMTP vocabulary trimming

#20700

Open