Portrait of Mingyu Sung
1-bit → 8-bit

Mingyu Sung

AI Research Engineer — Efficient GenAI Inference (Quantization · Caching · Acceleration)

Nota AI, Platform Team (Quantizer) · South Korea

52 citations/h-index 4/10 published, 4 preprints/16 merged PRs across 5 repos/as of 2026-11-09

Research interests
Efficient inference for any model family — LLMs, VLMs, VLAs, diffusion models — via quantization, KV-cache compression, cache-based acceleration, speculative decoding and kernel optimization
Education
Ph.D. in Artificial Intelligence, Kyungpook National University (2026)
Open source
Core contributor to the Nunchaku CUDA acceleration engine (ICLR 2025 Spotlight); 16 merged PRs across Nunchaku, Hugging Face Transformers, ByteDance DreamO and other projects

Experience

2026-02 — present
AI Research Engineer — Platform Team, Quantizer Part · Nota AI
  • Technical Lead for the FuriosaAI customer account (Apr 2026 – present): handle technical engagement with the customer and deliver a self-developed quantization tool.
  • Develop Nota's LLM/VLM quantization platform: post-training quantization (GPTQ, AWQ, SmoothQuant, rotation-based), KV-cache quantization, and low-bit export (NVFP4, GGUF).
  • Extend model support and evaluation across LLM, VLM, MoE, embedding and video-diffusion models.

Open source16 merged PRs

Nunchaku

nunchux-ai/nunchaku  ·  Core Contributor  ·  9 merged  ·  2025 – 2026-02
CUDA acceleration engine for 4-bit neural networks (ICLR 2025 Spotlight, 3.9k+ stars)
  • Cache optimizations — V2 FBCaching (#621), double FB cache, TeaCache batch processing (#601) — for 2–5x additional speedup
  • IP-Adapter (XLabs flux-ip-adapter-v2) support (#418)
  • Fixes: LoRA key mismatch (#557), ControlNet (#360, #452), Sana (#380), offload segfault (#440)
  • Forward-pass tensor-op cleanup (#491)

ComfyUI-nunchaku

nunchux-ai/ComfyUI-nunchaku  ·  Contributor  ·  2 merged  ·  2025
ComfyUI integration for the Nunchaku engine
  • IP-Adapter support for ComfyUI (#305)
  • Cache-mechanism file separation (#474)

DreamO

bytedance/DreamO  ·  Contributor  ·  1 merged  ·  2025
ByteDance's unified image-customization framework
  • Dynamic FBCache / DoubleFBCache support for Nunchaku engine integration (#104)

Transformers

huggingface/transformers  ·  Contributor  ·  3 merged  ·  2026
Hugging Face Transformers library
  • Fixed float16 overflow in Gemma4 vision pooler (#46277)
  • Fixed Gemma `sliding_window` being halved on every config save/reload, a silent accuracy regression for EmbeddingGemma (#47940)
  • Fixed `_init_weights` reading a quantized child's `weight`, which broke loading CLIP-family quantized checkpoints (#47921)

LoopQuant

Kishon-zzx/LoopQuant  ·  Contributor  ·  1 merged  ·  2026
Post-training quantization for looped language models (Ouro)
  • Fixed the LCTC teacher load resetting the ILPA per-loop activation-bit allocation, which ran the W4A6 configs at A4 in every loop; Ouro-2.6B W4A6 WikiText-2 perplexity 8.10 → 7.85 (#1)

Publications10 published

2026
A Bi-Objective Framework for Scene Text Super-Resolution: Balancing Legibility and FidelityTransactions on Machine Learning Research (TMLR) · M. Sung, S. Ham, K. Kim, I.-M. Kim, S. Yun, J.-M. Kang cited ×1 Accepted
Pose-Agnostic and Instance-Aware Test-Time Adaptation for Monocular Depth Estimation in IoTIEEE Internet of Things Journal · M. Sung, H. Choe, I.-M. Kim, S. Yun, J.-M. Kang Accepted
2025
H2-Cache: A Novel Hierarchical Dual-Stage Cache for High-Performance Acceleration of Generative Diffusion ModelsIEEE Open Journal of the Computer Society · M. Sung, I.-M. Kim, S. Yun, J.-M. Kang
A Novel VLM-Guided Diffusion Model for Remote Sensing Image Super-ResolutionIEEE Geoscience and Remote Sensing Letters · M. Sung, M.-G. Gong, S.-J. Ham, I.-M. Kim, S. Yun, J.-M. Kang cited ×2 Best Paper Award, KNU-EE Research Congress
DeCo-MeSC: Deep Compression-Based Memory-Constrained Split Computing Framework for Cooperative Inference of Neural NetworkIEEE Transactions on Vehicular Technology · M. Sung, V. Palakonda, I.-M. Kim, S. Yun, J.-M. Kang cited ×8
Generative Diffusion Model-Based Deep Learning Framework for Remaining Useful Life PredictionIEEE Internet of Things Journal · S. Ha*, M. Sung*, F. Saeed, S. Yun, I.-M. Kim, J.-M. Kang cited ×12 Co-first author
2024
Entropy-based sampling for efficient training of deep learning on CNC machining datasetElectronics Letters · M. Sung, C. Park, S. Ha, M. Ha, H. Lee, J. Kim, J.-M. Kang cited ×1
2023
Probabilistic Classification Method of Spiking Neural Network Based on Multi-Labeling of NeuronsMathematics · M. Sung, J. Kim, J.-M. Kang cited ×1
2021
Training and Inference using Approximate Floating-Point Arithmetic for Energy Efficient Spiking Neural Network ProcessorsICEIC · M. Kwak, J. Lee, H. Seo, M. Sung, Y. Kim cited ×6
2020
Training spiking neural networks with an adaptive leaky integrate-and-fire neuronIEEE ICCE-Asia · M. Sung, Y. Kim cited ×15

Preprints4 total

2026
RECAP-KV: Single-Prefill Approximation of Reconstruction Importance for Reusable KV Cache EvictionOpenReview · M. Sung et al.
2025
Memory- and Latency-Constrained Inference of Large Language Models via Adaptive Split ComputingarXiv:2511.04002 · M. Sung, V. Palakonda, S. Im, S. Moon, I.-M. Kim, S. Yun, J.-M. Kang
Why Should the Server Do It All?: A Scalable, Versatile, and Model-Agnostic Framework for Server-Light DNN Inference over Massively Distributed Clients via Training-Free Intermediate Feature CompressionarXiv:2511.11608 · M. Sung, S. Im, D. Bang, I.-M. Kim, S. Yun, J.-M. Kang
Range Asymmetric Numeral Systems-Based Lightweight Intermediate Feature Compression for Split Computing of Deep Neural NetworksarXiv:2511.11664 · M. Sung, S. Im, V. Palakonda, J.-M. Kang

Projects

2026 · Quantization Engineer
K-EXAONE 236B MoE Optimization (Nota AI × LG AI Research × FuriosaAI)
Optimized the 236B-parameter MoE model for FuriosaAI's data-center NPU via targeted precision analysis of degradation-prone sections — ~71% size reduction with ~99.2% accuracy retention (GPQA 79.80 / IFBench 68.98 / AIME25 88.57).
2024 · Lead Developer
LLM Inference Optimization (Samsung Challenge)
Optimized pipeline for Phi-3 using Torch-TensorRT conversion and memory-aware dynamic batching to accelerate LLM inference.
2023 · Lead Developer
Advanced CNC Machine Tool Diagnosis & Prognosis (KERI)
Deep learning models for tool wear classification, anomaly detection, and RUL prognosis.
2023 · AI Optimization Engineer
AI-based Intelligent CCTV (ABB Project)
Optimized an LSTR behavior-detection model for edge devices via Knowledge Distillation (TransKD) and Deep Compression — 25x model compression.
2022 · Participating Researcher
Global Basic Research Laboratory (NRF Project)
Deep learning-based channel estimation (CNN denoising, self-supervised learning) for IRS-aided communication systems.
2022 · Lead Developer
ML-based Sensor Data Analysis (JS System Project)
ML-based sensor feature-importance extraction and model performance optimization.

Education

2021-08 — 2026-02
Ph.D., Artificial Intelligence · Kyungpook National University (KNU)
Thesis: Fast and Trustworthy Super-Resolution with VLM-Guided Diffusion Model
2019-09 — 2021-08
M.S., Computer Science · Kyungpook National University (KNU)
Thesis: Probabilistic Classification Method of Spiking Neural Network Based on Multi-Labeling of Neurons
2013-03 — 2019-08
B.S., Computer Science · Kyungpook National University (KNU)

Awards

2025
Best Paper Award · KNU-EE Research Congress
A Novel VLM-Guided Diffusion Model for Remote Sensing Image Super-Resolution

Skills

Research
model quantizationKV-cache compressiondiffusion/LLM inference accelerationsplit computingsuper-resolutionlong-context evaluation
Programming
PythonC++CUDA CMATLAB
Frameworks & Tools
PyTorchHuggingFace TransformersTensorRTTritonflash-attnDockerGitLaTeX
Methodology
preregistrationclaims ledgerfreeze discipline