I patched llama.cpp to gain 20% prompt processing TPS. Help me make a PR

A developer patched llama.cpp to improve prompt processing throughput by 20% when using Multi-Token Prediction (MTP) on GPU, recovering performance lost to MTP overhead. The proof-of-concept processes only the output row instead of the entire ubatch for the last layer MoE FFN, restoring prompt processing TPS to non-MTP levels while retaining most MTP benefits for token generation. The developer seeks collaborators to open a pull request, as the code is AI-generated and cannot be submitted directly due to llama.cpp's contribution policy.

I've been running Qwen3.6-35B-A3B locally on llama.cpp and noticed that prompt processing throughput gets too low with MTP. I got nerd-sniped. What started as curiosity turned into a two-week rabbit hole of experiments and ended with a PoC that fully recovers the MTP PP overhead on GPU, above any expectation I had. TL;DR: instead of processing the last layer MoE FFN for the entire ubatch tokens usually 512-2048 tokens , this PoC processes only the output row usually 1 token during prefill . The result is PP TPS is back to the same as with MTP disabled, keeping most of MTP's benefits to TG TPS, even with a slight drop in draft acceptance rate in one of the benchs. I'm not opening a PR to llama.cpp because this is AI-generated code, which goes against their contribution policy, which I support. If you know C++ and llama.cpp internals, I invite to work together with me to open a PR with a more mature implementation. Comments URL: https://news.ycombinator.com/item?id=48700782 https://news.ycombinator.com/item?id=48700782 Points: 2 Comments: 1