focus-llama: a llama.cpp fork implementing Declarative Attention (arXiv:2609.02737)

A developer has forked llama.cpp to implement Declarative Attention, a technique from Google DeepMind and KAIST AI (arXiv:2609.02737). This approach allows the model to declare necessary context chunks in its output, which the engine then uses to restrict attention for subsequent tokens. The paper claims this method can reduce overall decode time to 0.71x for Gemma and 0.77x for Qwen compared to vanilla vLLM, without requiring additional training or a scorer. The developer is seeking feedback on the fork.