llama.cpp Adds MTP Support

MTP speculative decoding has been merged into llama.cpp mainline. PR #22673 (commit 4f13cb7) was merged on May 16, and it's enabled with --spec-type draft-mtp --spec-draft-n-max N. With the same seed and temperature, it produces byte-identical output compared to the baseline run.

Speed improvements were confirmed in single-stream chat benchmarks for Qwen3.6 27B.

• Strix Halo (Framework Desktop, ROCm 7.0.2): Q4_K_M 11.7 → 21.2 tok/s (1.81x), Q8_0 7.4 → 18.1 tok/s (2.44x)
• Single RTX 3090 (CUDA 12.9, driver 590.26): Q4_K_M 38.7 → 59.5 tok/s (1.54x, n=2)
• Dual RTX 3090 layer-split: Q8_0 25.7 → 55.9 tok/s (2.17x, n=3)

For Qwen3.6 35B-A3B MoE, the improvement margin was smaller.

• Strix Halo: 49.5 → 69.4 tok/s (1.40x)
• RTX 3090: 120.0 → 148.3 tok/s (1.24x)

The author explained that because MoE only runs about 3B/35B parameters per token, the absolute gain from speculative decoding is reduced. They also corrected that the earlier 3090 benchmarks were measured low due to a 200W cap, and after re-benchmarking at 350W/450W, confirmed +70%~+113% improvements on dense 27~32B models. The benchmark page now also includes prompt-processing tok/s and prompt-token columns.