@RadixAttention

Trellis introduced RadixAttention, a radix-tree-based KV cache designed to accelerate the prefill phase of LLM inference for chat and agentic sessions. The system stores shared string prefixes compact…