About Me
I build machine learning systems for large language models, focusing on efficient inference and agentic serving, speculative decoding, GPU kernel optimization, and post-training infrastructure.
I am an ML systems researcher at Carnegie Mellon's Catalyst Lab and Parallel Data Lab, advised by Zhihao Jia. I spent two years at Snowflake AI Research, working with Samyam Rajbhandari and Aurick Qiao. I am completing my Ph.D. in Computer Science at Carnegie Mellon University.
Highlights include: SpecInfer, whose tree-based verification is used by EAGLE-3, Medusa, and REST in vLLM and SGLang; SuffixDecoding, productionized at Snowflake for up to 5.3× agentic serving speedup; and FlexLLM, for up to 6.8× higher fine-tuning throughput.
Education
-
Carnegie Mellon UniversityPhD & MS in Computer ScienceDec. 2026 -
Tsinghua UniversityMS in Advanced ComputingJune 2023 -
Harvard UniversityBS in Electrical EngineeringMay 2021
Experience
-
Snowflake AI ResearchML Systems Research InternMay 2024 - Aug. 2026 -
CMU Catalyst GroupML Systems ResearcherAug. 2022 - Dec. 2026 -
Whist TechnologiesSoftware EngineerSep. 2021 - Aug. 2022
Service
Reviewer for ICLR '27, NeurIPS '26, ICML '26, MLSys '26, TMLR '26, ICLR '25, SIGMOD '24 (AEC), SOSP '24 (AEC)
Research (all publications )
LLM Inference & Serving
Fast, cost-efficient LLM serving under strict SLOs, including speculative decoding that accelerates inference with no loss in output quality.
- SuffixDecoding (NeurIPS 2025 🌟, US Patent 2026 📜)
- SpecInfer (ASPLOS 2024 🔥)
- AdaServe (EuroSys 2026)
- SpecReason (NeurIPS 2025)
- OWL (arXiv 2025)
LLM Fine-Tuning
Memory-efficient fine-tuning and token-level co-serving of inference and training on shared GPUs.
GPU Kernels & Compilation
Generating, compiling, and orchestrating high-performance GPU kernels for ML workloads.
- FastKernels (arXiv 2026)
- Event Tensor (MLSys 2026)
- Korch (ASPLOS 2024)
Networking & Telemetry
Line-rate network telemetry collection that bypasses the CPU on the data path.
- Direct Telemetry Access (SIGCOMM 2023)
- Zero-CPU Collection (HotNets 2021)
🏆 Outstanding Paper ·
🌟 Spotlight ·
🔥 750+ citations ·
📜 Patent
Honors & Awards
Silver Reviewer Award, ICML
2026
Top 1% Most-Cited Authors in Computer Science (OpenAlex, 5-year)
2025
Spotlight Presentation (Top 3.5%), NeurIPS
2025
Outstanding Paper Award (Top 0.5%), ACL
2024
Full-Ride Merit Scholarship, Tsinghua University
2021
Cum Laude with High Honors, Harvard University
2021