Carnegie Mellon University · Catalyst Group
Building efficient ML systems
and embodied models.
I'm a first-year PhD student at Carnegie Mellon University, advised by Tianqi Chen and Zhihao Jia. I received my B.S. in computer science from CMU. I work on embodied models, agentic GPU programming, LLM inference, and storage systems. I also collaborate with Juncheng Yang at Harvard SEAS on cache replacement and resilient routing for LLM inference.
Feel free to email me if you’d like to chat about research or explore a collaboration.
Email yiyanz [at] andrew [dot] cmu [dot] edu
GPU programming, language models, and caching.
An open-source environment that helps AI agents write, debug, and optimize GPU kernels. It combines the TIRx compiler foundation with a kernel knowledge base, analysis tools, and a benchmark server.
FlashInfer-Bench evaluates AI-generated GPU kernels on real LLM serving workloads and makes them easy to deploy. It checks correctness and performance, then enables kernel replacement in FlashInfer, SGLang, and vLLM.
A cache replacement algorithm designed for metadata in VMware vSAN. It distinguishes short bursts of access from reuse over time, helping the cache retain useful blocks and reduce misses with low overhead.
WebLLM runs large language models directly in the browser using local hardware. WebLLM Assistant builds on it to bring AI assistance to Overleaf and Google Workspace.
We train a language model to identify a text’s source by predicting a label as the next token, without a separate classification layer. We also introduce OpenLLMText, a dataset of about 340,000 human-written and model-generated samples.
Clock2Q+ uses three queues and a correlation window to prevent closely spaced accesses from being mistaken for lasting reuse. On the evaluated metadata traces, it achieves up to a 28.5% lower miss ratio than S3-FIFO. It has low CPU and memory overhead, scales across multiple CPUs, and has been deployed in VMware vSAN and VDFS.
FlashInfer-Bench brings kernel generation, benchmarking, and deployment into one workflow. Its shared trace format describes kernel definitions, workloads, implementations, and evaluation results. A dataset of real serving traces and a public leaderboard support comparisons across agents, while a kernel substitution mechanism lets inference engines use the best-performing implementations.
WebLLM is an open-source JavaScript framework for running LLMs entirely in the browser. It provides an OpenAI-style API, uses WebGPU for GPU acceleration and WebAssembly for CPU computation, and compiles optimized kernels with MLC-LLM and Apache TVM. Local execution lets applications run inference without sending prompts to a server.
This work treats text-source identification as next-token prediction. We fine-tune T5 to predict source labels without adding a classification layer and compare it with classification based on hidden states. We also introduce OpenLLMText, with about 340,000 samples from humans, GPT-3.5, PaLM, LLaMA, and GPT-2.
Open-source books on GPU programming for ML systems.
A practical guide to helping AI agents develop and optimize GPU kernels through compiler tools, a knowledge base, and performance feedback. Includes TIRx Harness and hands-on optimization examples.
A guide to GPU architecture and kernel optimization, from the TIRx programming model to GEMM and FlashAttention implementations. The examples focus on NVIDIA Blackwell.
Teaching Assistant · Graduate course
Machine Learning Department
Teaching Assistant · Undergraduate course
Department of Mathematical Sciences