Looking for a job

weiqi Fan

@weiqifan

AI infrastructure intern at Shanghai Innovation Institute who aligned distributed LLM training and inference and reduced MoE rollout from ~14,000 s to ~300 s.

China

Message

At Shanghai Innovation Institute's Systems Research Center, I contributed to an end-to-end Agentic RL stack and implemented the Miles-Polar Bridge. I also developed deterministic methods that aligned SGLang and Megatron computations, reducing Qwen3-4B and Qwen3-30B-A3B log-probability differences to zero.

On my JAX multi-GPU physics solver project, I built mappings for workload partitioning and implemented irregular many-to-many communication with mpi4jax. I also built a personal GPU performance tuning agent that reduced trials to a near-optimal configuration by ~35% in preliminary experiments versus Optuna alone.

Experience

Work history, roles, and key accomplishments

Current

AI Infrastructure Intern

Current

Shanghai Innovation Institute

Jun 2026 - Present (5 months)

Contributed to an end-to-end Agentic RL stack integrating Miles Training, Polar Rollout, Agent Fleet Runtime, SGLang, and Megatron. Implemented the Miles-Polar Bridge, aligned the Trajectory-to-Sample contract, and developed AgentFleetHarness with Agent Fleet/Harbor API support.

Sglang NVIDIA Megatron Agentic RL Distributed Training Python Graphics Performance Debugging

Education

Degrees, certifications, and relevant coursework

Sichuan University

Master of Science, Computer Science and Technology

2024 -

Pursuing a Master of Science in Computer Science and Technology, with research at the National Key Laboratory of Fundamental Algorithms and Models for Engineering Numerical Simulation under Prof. Xiaogang Deng.