The SyFI Lab at the University of Washington builds efficient and resilient infrastructure for the future of AI. As applications grow more complex, we bridge the gap between next-gen models and heterogeneous hardware through cross-stack innovation, delivering scalable, open-source systems validated by industrial partners.
Our research targets three key areas:
- Efficient AI: Optimizing algorithms and systems to maximize performance for training and inference.
- Flexible AI: Architecting systems that seamlessly adapt to diverse tasks, strategies, and model structures.
- Resilient AI: Ensuring AI system reliability at scale while leveraging AI to improve infrastructure robustness.
News
Publications
Ekka: Automated Diagnosis of Silent Errors in LLM Inference
Yile Gu, Zhen Zhang, Shaowei Zhu, Xinwei Fu, Jun Wu, Yida Wang, Baris Kasikci — International Conference on Machine Learning (ICML) (2026)
TraceLab: Characterizing Coding Agent Workloads for LLM Serving
Kan Zhu, Mathew Jacob, Chenxi Ma, Yi Pan, Stephanie Wang, Arvind Krishnamurthy, Baris Kasikci — (2026)
Piper: A Programmable Distributed Training System
Megan Frisella, Shubham Tiwari, Andy Ruan, Yi Pan, Parker Gustafson, Mat Jacob, Gilbert Bernstein, Stephanie Wang — (2026)
Blog Posts
Ekka: Automated Diagnosis of Silent Errors in LLM Inference
June 29, 2026
We present Ekka, an automated system that diagnoses silent errors in LLM serving frameworks via differential debugging: aligning and comparing a buggy framework's intermediate states against a trusted reference to pinpoint the root cause. Ekka reaches 80% pass@1 accuracy on 17 real-world vLLM and SGLang bugs and has already found 4 new ones confirmed by developers.TraceLab: Characterizing Coding Agent Workloads for LLM Serving
June 25, 2026
As coding agents become a major LLM application, serving them efficiently is an open systems problem. We release the SyFI coding trace — ~4,300 real sessions and 55B tokens for performance modeling from our daily Claude Code and Codex use — and TraceLab, an open pipeline to collect, sanitize, analyze, and replay your own coding agent traces.M*: A Modular, Extensible, Serving System for Multimodal Models
June 19, 2026
Today's models no longer fit the mold of autoregressive token generation. M* treats composite multimodal models as dataflow graphs and requests as Walks on those graphs, serving image, audio, video, and robot-action models at or above the performance of specialized engines.Talks
Enabling ASIC AI Chips for Cryptography Primitives
June 12, 2026 Jianming Tong — Georgia Tech
ML for Systems in the Real World: Challenges Beyond Training the Models
May 29, 2026 Jiayi (Jane) Chen — The University of Texas at Austin
Extreme Codesign for Efficient AI: Algorithms, Software, and Hardware
May 22, 2026 Mohamed Abdelfattah — Cornell University