Chenrui Gao

I’m Chenrui, an incoming M.S. student in Computer Science at UT Austin. I did my B.S.E. in Computer Engineering at the University of Michigan as part of a dual-degree program with Shanghai Jiao Tong University.

I work on LLM training and inference systems, mostly as an upstream contributor to open-source RL post-training and serving frameworks:

• verl-omni (committer) — RL post-training for diffusion and omni models: brought on-policy distillation to diffusion RL, from the teacher-anchored losses through multi-teacher routing, standalone teacher pools, and async teacher scheduling on the v1 trainer; also worked on rollout–training numerical consistency, profiling, and adapter correctness.
• sglang-omni — inference performance for Qwen3-Omni: batching the multimodal prefill path and removing host syncs on the way to CUDA-graph coverage.
• vllm-omni — parallelism and acceleration for video diffusion models: sequence, tensor, and CFG parallelism plus caching and offload for SANA-Video, regional torch.compile for MAGI-2 Preview, and attention-backend and offload bug fixes.

Before that I did robotics research at CMU’s ARCS Lab with Prof. Jiaoyang Li, where I built the SIMD-vectorized core of VAMP-MR (IROS 2026), and at UC Irvine with Prof. Sven Koenig on topological multi-agent pathfinding. Full details are on the cv page.