Zhihao Jia
@JiaZhihao
Followers
3K
Following
570
Media
22
Statuses
195
Assistant professor of Computer Science at Carnegie Mellon University. Research on systems and machine learning.
Joined August 2012
One of the best ways to reduce LLM latency is by fusing all computation and communication into a single GPU megakernel. But writing megakernels by hand is extremely hard. 🚀Introducing Mirage Persistent Kernel (MPK), a compiler that automatically transforms LLMs into optimized
12
119
767
RT @WentaoGuo7: 🦆🚀QuACK🦆🚀: new SOL mem-bound kernel library without a single line of CUDA C++ all straight in Python thanks to CuTe-DSL. On….
0
61
0
RT @reyna_abhyankar: Computer-Use Agents (CUAs) are improving every day but take up to tens of minutes to complete simple tasks. We built O….
0
2
0
RT @FrancisYan_: 🚀 [OSDI ’25, Tue 11:10am].How do you “divide and conquer” large-scale resource allocation problems like GPU cluster schedu….
0
5
0
RT @NovaSkyAI: ✨Release: We upgraded SkyRL into a highly-modular, performant RL framework for training LLMs. We prioritized modularity—easi….
0
43
0
RT @anjiangw: We introduce CodeARC, a new benchmark for evaluating LLMs’ inductive reasoning. Agents must synthesize functions from I/O exa….
0
28
0
RT @JeffDean: Mark your calendars for #MLSys2026 in May, 2026 in Seattle. Submission deadline for papers is Oct 30 this year.
0
14
0
RT @tqchenml: #MLSys2026 will be led by the general chair @luisceze and PC chairs @JiaZhihao and @achowdhery. The conference will be held i….
0
11
0
RT @JiaZhihao: 📢Exciting updates from #MLSys2025! All session recordings are now available and free to watch at We….
0
29
0
📢Exciting updates from #MLSys2025! All session recordings are now available and free to watch at We’re also thrilled to announce that #MLSys2026 will be held in Seattle next May—submissions open next month with a deadline of Oct 30. We look forward to
2
29
103
✨MPK is very easy to use. You can compile your LLMs into a high-performance megakernel for fast inference with just a few dozen lines of Python, no CUDA required.
0
0
28
⚡MPK pushes LLM inference latency towards hardware limits. On a single A100-40GB GPU, MPK reduces Qwen3-8B per-token latency from 14.5ms (vLLM/SGLang) to 12.5ms—approaching the theoretical lower bound of 10ms based on memory bandwidth.
1
1
33
RT @BeidiChen: Say hello to Multiverse — the Everything Everywhere All At Once of generative modeling. 💥 Lossless, adaptive, and gloriousl….
0
20
0
RT @tqchenml: Check out our work on parallel reasoning 🧠; We bring an AI-assisted curator that identifies parallel paths in sequential trac….
0
15
0
RT @yi_xin_dong: @databricks 's Agent Bricks is powered by XGrammar for structured generation, and achieves high quality and efficiency. It….
0
4
0
RT @matei_zaharia: Excited to launch Agent Bricks, a new way to build auto-optimized agents on your tasks. Agent Bricks uniquely takes a *d….
0
45
0
RT @SCSatCMU: A collaborative research project led in part by SCS researchers Tianqi Chen and Ruihang Lai received a Best Paper award at th….
0
2
0