RIS-Kernel: A Model-Agnostic Architecture for Long-Context LLM Inference via Sparse Attention
The RIS-Kernel engine solves large language model context limits by using model-agnostic sparse attention to reduce complexity from O(N^2) to O(N log N). It achieves this by injecting runtime sparsity into unmodified models using two sampling modes: Stochastic for global coverage and Structural for local preservation. This makes long-context LLM inference possible on standard CPU hardware without needing expensive GPUs.