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
A Self-Defining Operator for Distributed Applications
Shihang Li, Thomas Anderson, Ratul Mahajan, Simon Peter, Luke Zettlemoyer — (2026)
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)
Blog Posts
Murmur: Efficient Long-Context ASR Inference with Hierarchical Chunking
November 13, 2026
Murmur speeds up long-context ASR models such as VibeVoice-ASR and MOSS-Transcribe-Diarize by 2.5× on average, end to end including diarization, by decoding ~5-minute chunks in parallel and drafting tokens from a small model's transcript of 30 s windows. It also lowers speaker-attributed error compared with WhisperX on every dataset we tested.VibeSys on Trainium: Agents Build and Tune Qwen3.8 Serving Engines on AWS Trainium2
November 04, 2026
We used VibeSys to build serving engines for two new hybrid Qwen3.8 models on AWS Trainium2, which vLLM Neuron does not support. On one Trainium2 chip, agents wrote a Qwen3.8-27B engine from scratch and raised its throughput from 195 to 861 tok/s, starting from a basic PyTorch implementation. On a 16-chip node, VibeSys guided by ServingStudio made a simple Qwen3.8-Flash-Next engine 3.1× faster.VibeSys Builds a Qwen3.5-397B Engine on MI300A, 2.3× Tuned SGLang
October 08, 2026
VibeSys's agents built a serving engine from scratch for Qwen3.5-397B-A17B on four AMD MI300As, taking it from 12.8 to 2,242 tok/s, 2.33× the goodput of an SGLang that we first had to port and tune.Talks
Context Language Models
October 23, 2026 Rulin Shao — University of Washington
Unlocking the Full Potential of TorchDynamo: Accelerating, Comparing, and Debugging ML Systems
October 16, 2026 Yi Pan — UC Berkeley
How not to build another Megatron, for both humans and agents
October 9, 2026 Lequn Chen — Perplexity AI