My name is Xueshen Liu (刘学深). I am a third-year Ph.D. candidate in Computer Science and Engineering Division at the University of Michigan, advised by Prof. Z. Morley Mao. My research interests focus on building systems bridging the gap between software and hardware. I am currently working on accelerating large language models (LLMs) through efficient system-algorithm co-design.
Education:
- University of Michigan (2022 - Present)
Ph.D of Computer Science & Engineering (CSE) - University of Michigan (2020 - 2022)
Bachelor of Computer Science (CS) - Shanghai Jiao Tong University (2018 - 2022)
Bachelor of Electric & Computer Engineering (ECE)
Projects:
- Low-cost Distributed MoE Training (April 2024 - Present):
On-going project to reduce the cost of training MoE models by 50%. - Computation-awared KV Cache (CAKE) Loader (Sept. 2024 - Present):
CAKE is a novel KV cache loader that reduces LLM prefill latency on long-context through a bidirectional KV cache generation strategy overlapping computation and I/O transfer. - Learn To be Efficienct (LTE) (Mar. 2024 - Dec. 2024):
Learn-To-be-Efficient (LTE) trains LLMs to activate fewer neurons through structured sparsity while maintaining performance, achieving linear speedup with our custom CUDA kernel. - Minimap2-gigabase (May 2022 - April 2024):
Minimap2-gigabase (mm2-gb) is based on minimap2-v2.24 with GPU accelerated chaining kernel for high throughput accurate mapping of ultra-long reads. - IMAGician (Jan. 2022 - May 2022):
IMAGician is an Android App based on Image Steganography Technology. It embeds invisible watermarks in images to protect copyright of creators. Check our demo video on Youtube.
Experience:
- Reviewer for ICLR’25, COLING’25 (Oct. 2024)
- Graduate Student Instructor for CSE 589 (Advanced Computer Networks) in UMich (Sept. 2024 - Dec. 2024)
- Connected Autonomous Vehicle (CAV) Research Intern at General Motors (May 2024 - Aug. 2024)