Zheng Cai (@zigzagcai) profile picture

ML System

Shanghai
Joined October 2026
Photo by Zheng Cai (@zigzagcai)
I built agent-driven framework with LoopX as workflow engine to automatically optimize inference performance on vLLM. As a case, automatically discovered sequence parallelism for sparse indexer in GLM-5.3-Flash, resulting in 32.6% throughput increase under 1M long context length.
7
2
1
57
15,397
Through continuous iteration, the agent discovered an efficient sequence-parallel implementation for sparse indexer; this boosted inference throughput of GLM-5.3-Flash by 32.6% at a 1M-token context length, and the method is applicable to other models utilizing sparse attention.
3
588
This framework enables agents to autonomously perform profiling, explore research papers and codebases, formulate hypotheses, write code to conduct experiments, and measure correctness and speedup.
3
806