kv-cache-efficiency

This paper introduces Scratchpad Patching, a technique for tokenizer-free language models that decouples compute from patch size by dynamically refreshing context within patches to reduce patch lag.