Open to opportunities

Namgyu Youn

@namgyuyoun

Software engineer focused on LLM inference and quantization, delivering memory-efficient serving and optimized inference performance.

South Korea

Message

What I'm looking for

I’m looking for a role where I can drive LLM inference and quantization work—focusing on memory-efficient serving, kernel-level performance, and real-world deployment pipelines—while continuing to contribute to open-source runtimes like vLLM and torchao.

I’m a software engineer focused on LLM inference and quantization, with experience in memory-efficient serving and inference performance optimization. I’ve identified and removed redundancy in Qwen3-14B ONNX export graphs, then built an INT8 TensorRT deployment pipeline using QDQ calibration to achieve measurable latency and memory gains.

I contribute to open-source across vLLM quantization dispatch, KV cache scheduling, and runtime modules, and have 25+ merged PRs to torchao (maintained by PyTorch). I’ve also upstreamed quantization methods and benchmark infrastructure, designed low-bit tensor types for memory-efficient LLM/VLM inference, and implemented low-bit MatMul kernels in Triton for practical throughput improvements.

Experience

Work history, roles, and key accomplishments

Current

vLLM Contributor

Current

vLLM

May 2026 - Present (2 months)

Contributing across quantization dispatch (NVFP4/RoPE), KV cache scheduling, and runtime modules in vLLM.

VLLM Quantization NVFP4 ROPEs

Deep Learning Engineer

CLIKA

Feb 2026 - Mar 2026 (1 month)

Identified redundancy in a Qwen3-14B ONNX export graph and applied graph fusion (Attention, LayerNorm) to reduce node count by 21%. Built an INT8 TensorRT deployment pipeline via QDQ calibration, achieving 36% lower latency and 41% lower memory vs FP16.

Onnx Graph Fusion TensorRT QDQ Calibration NVIDIA RTX 5090 Inference Optimization

Efficient AI Inference Lab

AerLabs

Sep 2025 - Jan 2026 (4 months)

Presented LLM inference serving research including AWQ (MLSys’24), QServe (MLSys’25), and Flash Attention. Discussed kernel-level implementations for efficient inference.

LLM Inference Flash Attention Research Presentation

torchao (ao) Contributor

PyTorch Foundation (Meta)

Jul 2025 - Jan 2026 (6 months)

Upstreamed PTQ benchmarking infrastructure (lm-eval, vLLM) and quantization methods (AWQ/SmoothQuant), and built torchao’s first end-to-end GPU memory benchmark. Designed INT8/MXFP8 tensor subclasses for memory-efficient LLM/VLM inference and implemented low-bit MatMul kernels in Triton (INT4/INT8/MXFP8).

TorchAO PyTorch VLLM SmoothQuant GPU Memory Benchmarking INT8 LLM Inference Triton Low Bit MatMul INT4

Optimizers

Meta Research

Mar 2025 - May 2025 (2 months)

Improved QAT accuracy by implementing a Hadamard rotation kernel (C++) to reduce quantization error. Profiled distributed Shampoo using PyTorch Profiler and identified root inverse and eigenvector computations as major overhead.

QAT Hadamard Rotation Kernel C++Distributed Shampoo PyTorch Profiler

Undergraduate Research Assistant

Sogang University

Nov 2022 - Jun 2023 (7 months)

Built a benchmarking pipeline for YOLO- and DETR-based biological object detection models, evaluating mAP, latency, and throughput to guide model selection. Exported the selected DETR model to ONNX and validated inference correctness for deployment.

YOLO Biological Object Detection Map Evaluation Throughput Onnx