llama.cpp Adds Beta Support for MTP

llama.cpp has added beta support for MTP (Multi Token Prediction). MTP models are automatically loaded from the same GGUF and use a separate context and KV cache.

• To prevent hidden features from not being passed across multiple ubatch calls, a separate hook consumed by MTP was added after each ubatch.
• In the Qwen3.6 27B benchmark, the --spec-type mtp --spec-draft-n-max 3 setting recorded an average acceptance rate of 72.18%, and total wall time dropped from 201.07 seconds → 83.8 seconds.
• The baseline benchmark was 7.0–7.7 tok/s, and with MTP applied it rose to 13.9–21.6 tok/s.

The GGUF for MTP and the changes to convert_hf_to_gguf.py have also been released, allowing immediate experimentation in llama-server.