CV

Curriculum vitae.

Contact Information

Name Omin Kwon
Email om0127@snu.ac.kr
Location Seoul, South Korea
Website https://omin-kwon.github.io
LinkedIn ominkwon

Research Interest

  • Building efficient and scalable ML systems through algorithm-system co-design.

Education

  • Sept 2025 - present

    Seoul, South Korea

    M.S.
    Seoul National University
    Computer Science and Engineering
    • Advisor: Prof. Jae W. Lee, ARC Lab
    • GPA: 4.3/4.3
  • Mar 2019 - Aug 2025

    Seoul, South Korea

    B.S.
    Seoul National University
    Electrical and Computer Engineering (summa cum laude)
    • GPA: 4.21/4.3 (upper-division), 3.95/4.3 (overall)
    • Ranked 1st in Machine Learning Fundamentals, Computer Architecture, and Computer Networks.
    • Leave of absence for service in the Republic of Korea Air Force (May 2021-Feb 2023)
  • Mar 2017 - Mar 2019

    Seoul, South Korea

    Hansung Science High School
    Early Graduation for Academic Excellence

Publications

  • 2026
    SketchSSM: Write to the Full State, Read from a Compact Sketch
    Omin Kwon, JoongWon Shin, Minseo Kim, Kurt Keutzer, Sehoon Kim, Jae W. Lee
    NeurIPS 2026 Long-Context Foundation Model Workshop (Oral) Under review at ICLR 2027 (Avg. Score: 3.0/4; 1,650/40,156 = 4.1%)
  • 2026
    HERALD: High Throughput Block Diffusion LLM Inference via CPU-GPU Cooperative KV Cache Retrieval
    Omin Kwon, Doyeon Kim, Jongseok Park, Seung Yul Lee, Sehoon Kim, Ion Stoica, Jae W. Lee
    Under review at MLSys 2027
  • 2026
    Triage: Skip, Skim, or Refine Attention Tiles in Video Diffusion Transformers
    Yeonjae Kim, Omin Kwon, Seong Hoon Seo, Jae W. Lee
    Under Review
  • 2026
    MAGE: All-[MASK] Block Already Knows Where to Look in Block Diffusion LLM
    Omin Kwon, Yeonjae Kim, Doyeon Kim, Minseo Kim, Yeonhong Park, Jae W. Lee
    NeurIPS 2026 ICML 2026 Adaptive Foundation Model Workshop (Oral, #text(fill: maroon)[Best Paper Award] (1/218 = 0.46%))
  • 2025
    STAR: Improving Lifetime and Performance of High-Capacity Modern SSDs Using State-Aware Randomizer
    Omin Kwon, Kyungjun Oh, Jaeyong Lee, Myungsuk Kim, Jihong Kim
    ICCAD 2025 (Oral)

    U.S. patent application with SK hynix: No. 19/569899, filed March 17, 2026.

  • 2025
    NestedFP: High-Performance, Memory-Efficient Dual-Precision Floating Point Support for LLMs
    Haeun Lee, Omin Kwon, Yeonhong Park, Jae W. Lee
    NeurIPS 2025
  • 2025
    AiDE: Attention-FFN Disaggregated Execution for Cost-Effective LLM Decoding on CXL-PNM
    KyungSoo Kim, Omin Kwon, Yeonhong Park, Jae W. Lee
    IEEE CAL

Research Experience

  • Seoul National University
    • Research visit to FAST Lab, UIUC (Mar-May 2026).
    • MAGE (Project Lead) · Advisor: Prof. Jae W. Lee - Designed a training-free sparse attention algorithm for block diffusion LLMs. - Developed an inference system integrated into SGLang with optimized CUDA kernels. - Achieved up to 6.82× end-to-end speedup at 128K context with near-lossless accuracy.
    • HERALD (Project Lead) · Advisor: Prof. Jae W. Lee - Designed a training-free sparse attention algorithm for block diffusion LLMs with KV offload. - Developed an inference system integrated into SGLang with a dual-stream CPU-GPU pipeline. - Achieved up to 3.70×, 2.59×, and 8.38× decode throughput over Dense, MAGE, and MAGE-Offload, respectively, with near-lossless accuracy at K=1024 across four block dLLMs.
    • SketchSSM (Project Lead) · Advisors: Prof. Jae W. Lee and Prof. Sehoon Kim - Designed a training-free linear attention algorithm that reduces state-read I/O. - Developed an inference system integrated into vLLM with optimized CUDA kernels. - Achieved up to 2.77× decode throughput over Standard vLLM with near-lossless accuracy.
  • Seoul National University
    • NestedFP · Advisor: Prof. Jae W. Lee - Integrated the NestedFP CUDA kernel into vLLM with a workload-aware precision-switching scheduler.
    • AiDE · Advisor: Prof. Jae W. Lee - Proposed workload-aware KV cache placement for CXL-PNM and developed an LLM simulator.
    • STAR (Project Lead) · Advisor: Prof. Jihong Kim - Proposed a 3D NAND randomization algorithm that extends SSD lifetime by up to 2.3×. - Designed STAR hardware and integrated it into the SSD controller. - Key contributor to a U.S. patent application with SK hynix on STAR.

Teaching Experience

  • Seoul National University
    {“company”=>”Seoul National University”, “position”=>”Teaching Assistant, Advanced Computer Architecture (4190.571)”, “location”=>”Seoul, South Korea”, “start_date”=>”2026-09”, “end_date”=>”present”}
  • Seoul National University
    • Selected as an official departmental tutor based on academic excellence.
    • Courses: Computer Organization and Design, Digital System Design, and Microelectronics.

Honors & Awards