Chenrui Gao

I’m Chenrui, an M.S. student in Computer Science at UT Austin, previously at the University of Michigan and Shanghai Jiao Tong University.

I work on LLM and diffusion systems: RL post-training frameworks and inference engines, almost all of it upstream. I am a committer of verl-omni and contribute to verl, vllm-omni and sglang-omni.

RL post-training

• On-policy distillation for diffusion RL · verl-omni
  Built end to end: distillation losses, a frozen-teacher runtime and multi-teacher routing. The student matches its RL teacher about 5× faster than training with GRPO.
• Async trainer throughput · verl-omni
  Async teacher scheduling cut the teacher wait from 3.6 s to 0.02 s, and hybrid rollout switching raised Wan2.2 video throughput by 22% on the same GPUs.
• Training–sampling mismatch · verl, verl-omni
  Score centering kept training stable with an fp8 sampler where plain policy gradient collapsed, and a mapping fix cut BAGEL’s log-prob drift 157×. In progress: bitwise on-policy RL for Qwen3-Omni, at zero mismatched tokens on the full model in a two-step closed-loop run.

Inference engines

• Video diffusion acceleration · vllm-omni
  Sequence parallelism for linear attention in SANA-Video, 1.80× on two GPUs and 3.57× with CFG parallelism on four, and regional torch.compile for MAGI-2. In progress: SANA-Video 2.0.
• Engine core · vllm-omni
  A keep-mode pause that quiets a busy diffusion replica without discarding finished work, plus fixes for silent sleep and wake-up failures, an orchestrator hang and a batch-mixing attention bug. In progress: prefix caching under preemption.
• Multimodal prefill · sglang-omni
  A batched embedding merge for Qwen3-Omni: host syncs fell from 96 to 0 per batch and the merge ran 2.1–4.1× faster.

Before this I did robotics research at CMU’s ARCS Lab (VAMP-MR, IROS 2026) and at UC Irvine. Details and numbers are on the cv page.