At StepFun, I work on post-training and RL infrastructure. I brought BF16/FP8 training and inference to bitwise parity, reducing actor and rollout logprob differences to zero, and optimized Step3.x deployment scheduling, CP parallelism, and KV offload.
Previously at Tencent WXG, I led projects spanning audio inference, multi-level KV caching, and LLM MaaS. My work reduced multi-turn audio first-token latency from 800 ms to 50 ms and helped raise compute-resource utilization from 50% to over 75%.
