Search papers, labs, and topics across Lattice.
Huazhong University of Science and Technology 搂 Xiaohongshu Inc
2
0
3
FlowBlock achieves up to 4.01脳 faster decoding and 77.1% lower latency in dLLMs by transforming block dependencies into scheduling resources through innovative wavefront-parallel techniques.
Transforming the KV cache from a monolithic structure into a dynamic, head-aware system could revolutionize LLM serving efficiency and scalability.
