About me

Welcome to Wei-Lin Chiang’s page!

• I am a CS PhD student in UC Berkeley Sky Computing Lab (previously RISElab), working with Prof. Ion Stoica.
• I obtained my bachelor’s and master’s degree from National Taiwan University under the supervision of Prof. Chih-Jen Lin.
• My research interests include scalable AI systems, cloud computing. Check out our intercloud broker system, SkyPilot, and FastChat, a chat LLM framework.
• I enjoy developing open-source ML software and I am always happy to learn how they are being used! Email me if you have questions or find our projects useful.
• More details can be found in my CV.

Projects

• SkyPilot (GitHub | Paper)
  SkyPilot is an open-source intercloud broker system for deploying AI workloads on any cloud; all major clouds (AWS/GCP/Azure) are supported; adopted by 10+ labs and organizations
• Vicuna (Blog | Demo | Weights)
  An open-source chatbot impressing GPT-4 with 90%* ChatGPT quality; Our demo has served 2 million requests; 400+ models on HuggingFace are based on Vicuna
• FastChat / Chatbot Arena (GitHub | Arena | leaderboard)
  An open platform for training, serving, and evaluating LLMs, powering Vicuna and Chatbot Arena; Chatbot Arena has collected 40K anonymous battles with human votes on 20+ LLMs
• LLM as a Judge (GitHub | Paper | Demo)
  LLM judges for chatbot evaluation with a multi-turn chat benchmark MT-Bench; Scalable, effective, and validated benchmark distinguishing 30+ chat LLMs (Leaderboard)
• Balsa (GitHub | Paper)
  Balsa is a ML-based query optimizer, learning to optimize SQL queries by trial-and-error using deep RL and sim-to-real learning
• Cluster-GCN (GitHub | Paper)
  Scalable training method for large (million-scale) and deep GCN Achieved state-of-the-art performance on public datasets (e.g., PPI, Reddit)
• Distributed LIBLINEAR (Code | Paper)
  Distributed extension of a widely-used linear classification package, LIBLINEAR
  Developed L1-regularized LR solver for solving large (billion-scale) tasks
• Multi-core LIBLINEAR (Code | Paper)
  Multi-core extension of a widely-used linear classification package, LIBLINEAR
  Developed efficient parallel algorithms for primal and dual solvers

Work Experience

• Intern@Amazon, Seattle (May. 2021 - Aug. 2021)
  Contrastive learning for information extraction on semi-structure webpages
• Intern@Google Research, Mountain View (Dec. 2018 - Mar. 2019)
  Efficient algorithms for training large and deep GCN models.
  Cluster-GCN paper, code
• Intern@Alibaba Group, Hangzhou (July 2017 - Sept. 2017)
  Distributed ML algorithms on Alibaba’s parameter server (KunPeng)
• Intern@Microsoft Research Asia, Beijing (Dec. 2016 - Feb. 2017)
  Distributed training for deep learning frameworks
• Intern@Microsoft, Redmond (July 2016 - Oct. 2016)
  Large-scale ML algorithms on Microsoft’s distributed platform (REEF)

Publications (Google Scholar Profile)

• Judging LLM-as-a-judge with MT-Bench and Chatbot Arena
  Lianmin Zheng^*, Wei-Lin Chiang^*, Sheng Ying^*, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhouhan Li, Dacheng Li, Eric Xing, Hao Zhang, Joseph Gonzalez, Ion Stoica (*equal contribution)
  In submission to NeurIPS 2023 Dataset and Benchmarks Track
• Vicuna: An Open-Source Chatbot Impressing GPT-4 with 90%^* ChatGPT Quality
  Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph Gonzalez, Ion Stoica, Eric Xing (alphabetical order)
  Blog post
• Optimizing Spot Instance Savings under Deadlines
  Zhanghao Wu, Wei-Lin Chiang, Zongheng Yang, Eric Friedman, Scott Shenker, Ion Stoica.
  In submission to NSDI 2024
• SkyPilot: An Intercloud Broker for Sky Computing
  Zongheng Yang, Zhanghao Wu, Michael Luo, Wei-Lin Chiang, Romil Bhardwaj, Woosuk Kwon, Siyuan Zhuang, Frank Sifei Luan, Gautam Mittal, Scott Shenker, Ion Stoica
  USENIX NSDI 2023
• Balsa: Learning a Query Optimizer Without Expert Demonstrations
  Zongheng Yang, Wei-Lin Chiang⁺, Sifei Luan⁺, Gautam Mittal, Michael Luo, Ion Stoica. (+ equal contribution)
  ACM SIGMOD 2022
• Manifold Identification for Ultimately Communication-Efficient Distributed Optimization
  Yu-Sheng Li, Wei-Lin Chiang, and Ching-pei Lee.
  International Conference on Machine Learning (ICML), 2020
• Cluster-GCN: An Efficient Algorithm for Training Deep and Large Graph Convolutional Networks [code, dataset (Amazon2M)]
  Wei-Lin Chiang, Xuanqing Liu, Si Si, Yang Li, Samy Bengio, and Cho-Jui Hsieh.
  ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (KDD), 2019 (Oral) slides, poster
• Preconditioned Conjugate Gradient Methods in Truncated Newton Frameworks for Large-scale Linear Classification [supplement & code. Implementation available in LIBLINEAR after version 2.20.]
  Chih-Yang Hsia, Wei-Lin Chiang, and Chih-Jen Lin.
  Asian Conference on Machine Learning (ACML), 2018 (Best paper award) slides, poster
• Limited-memory Common-directions Method for Distributed L1-regularized Linear Classification [supplement & code. Implementation available in Distributed LIBLINEAR.]
  Wei-Lin Chiang, Yu-Sheng Li, Ching-pei Lee, and Chih-Jen Lin.
  SIAM International Conference on Data Mining (SDM), 2018 slides, poster
• Parallel Dual Coordinate Descent Method for Large-scale Linear Classification in Multi-core Environments [supplement, code. Implementation available in Multi-core LIBLINEAR.]
  Wei-Lin Chiang, Mu-Chu Lee, and Chih-Jen Lin.
  ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (KDD), 2016 poster
• Fast Matrix-vector Multiplications for Large-scale Logistic Regression on Shared-memory Systems [supplement, code. Implementation available in Multi-core LIBLINEAR.]
  Mu-Chu Lee, Wei-Lin Chiang, and Chih-Jen Lin.
  IEEE International Conference on Data Mining (ICDM), 2015 slides