Deep Learning Systems · LLM Inference

Yangmin Li a.k.a. YAMY

Deep Learning Algorithm Engineer at NVIDIA, working on SGLang & TensorRT-LLM — pushing LLM inference fast, stable, and correct at scale on GB200 / GB300.

Previously CMU INI & Tongji CS. I care about the intersection of engineering, optimization, and research with real-world impact.

Yangmin Li

01 — About

Building the systems that serve LLMs.

I’m a Deep Learning Algorithm Engineer at NVIDIA on the TensorRT-LLM / Inference team, where I focus on making large language models run faster, cheaper, and correctly at scale — most recently on Blackwell-generation hardware (GB200 / GB300 NVL72). A lot of my day-to-day lives inside SGLang: Qwen3.5 / Qwen3-Next bring-up, DeepSeek-V4 performance, heterogeneous-TP KV transfer, DP-Attention, and long-context disaggregated serving.

On the side, I collaborate with LMSYS and UC Berkeley’s Sky Computing Lab on multi-LLM serving (Prism, kvcached), and with CMU’s AI Common Sense Lab under Deepak Pathak on physics-grounded reasoning (Sim2Reason) — turning physics simulators into infinite training environments for LLMs.

Before NVIDIA full-time, I was at CMU INI (M.S., 2023–2024), and Tongji University for my B.E. in CS. I’ve been lucky to publish first-author papers at CIKM and IEEE TAFFC along the way.

02 — Now

What I’m working on

Updated 2026

Helping bring up Qwen3.5 / Qwen3-Next in SGLang from the NVIDIA side — working on heterogeneous-TP KV transfer for GQA, Mamba state slice transfer in disaggregation, a GPU staging-buffer with dynamic ring allocator, and fused QKV-Z-BA reshape on prefill.

Working on DeepSeek-V4 / V4-Pro performance on GB200 / GB300 NVL72 alongside the team — disaggregated 1P1D recipes, mega-MoE serving knobs, sa-bench tokenizer support, and PP / TP / EP / DP sweeps to map throughput × latency × memory Pareto frontiers.

Chipping away at SGLang inference critical paths: DP-Attention with reduce_scatterv in place of all-reduce + dp_scatter, an adaptive queue-based prefill delayer, TBO MLP-sync fixes, and NIXL transport for SWA / NSA / Mamba state buffers.

Continuing to contribute to Prism & kvcached — tighter SGLang / vLLM integration, cluster-level deployment, and Kubernetes-friendly multi-model serving.

03 — Selected Work

Projects & contributions

SGLang · NVIDIA · 2026

Qwen3.5 / Qwen3-Next — disagg & hybrid-attention

Contributing from the NVIDIA side: heterogeneous-TP KV transfer for GQA, Mamba state slice transfer in disaggregation, a GPU staging-buffer with dynamic ring allocator, fused QKV-Z-BA prefill, GDN on SM100+. End-to-end Qwen3.5 disagg serving.

SGLang · NVIDIA · 2026

DeepSeek-V4 performance on GB200 / GB300

Working with the team on disaggregated 1P1D recipes for GB300, mega-MoE serving knobs, sa-bench tokenizer support, and PP / TP / EP / DP sweeps mapping throughput–latency–memory Pareto frontiers for production serving.

LMSYS × Berkeley Sky Lab · 2025–

Prism — Multi-LLM serving on shared GPUs

Two-tier SLO-aware scheduler (global KV-pressure placement + GPU-local priority admission). 50+ LLMs concurrently on shared infra, 2× cost reduction & 3.3× SLO attainment.

arXiv 2505.04021 · Multi-LLM/prism-research →
LMSYS × Berkeley Sky Lab · 2025–

kvcached — GPU virtual memory for KV cache

OS-style virtual memory for LLM KV caches. CUDA virtual-memory APIs decouple physical from virtual GPU memory, enabling elastic allocation and cross-model cache sharing. Integrated with SGLang & vLLM.

Apache 2.0 · ovg-project/kvcached →
CMU AI Common Sense Lab · 2024–

Sim2Reason — Solving Physics Olympiad via RL on simulators

Procedurally generate MuJoCo scenes & QA pairs through a DSL, then RL-finetune LLMs on synthetic supervision. +10–15pp on IPhO Mechanics across model scales (Qwen2.5-32B: 19.8%→25.2%). With Mihir Prabhudesai, Aryan Satpathy, Deepak Pathak, Katerina Fragkiadaki.

sim2reason.github.io →
CMU ADMIS Lab · TAFFC’25

CorMulT — Modality correlation-aware Transformer

Semi-supervised model with modality-correlation contrastive learning for audio / image / text fusion. Strong results on CMU-MOSEI. First-author, IEEE Transactions on Affective Computing 2025.

04 — Publications

Selected papers

IEEE TAFFC · 2025

CorMulT: A Semi-supervised Modality Correlation-aware Multimodal Transformer for Sentiment Analysis

Yangmin Li, et al. · IEEE Transactions on Affective Computing, 2025 · first author

CIKM · 2022

An Extreme Semi-Supervised Framework Based on Transformer for Network Intrusion Detection

Yangmin Li, et al. · ACM CIKM 2022 · first author · ACM DL

Preprint · 2025

Prism: Unleashing GPU Sharing for Cost-Efficient Multi-LLM Serving

With Sky Computing Lab & LMSYS · arXiv 2505.04021

Project · 2024–

Sim2Reason: Solving Physics Olympiad via Reinforcement Learning on Physics Simulators

With CMU AI Common Sense Lab · project page

05 — Experience

Where I’ve been

  1. Jan 2025 — Present

    NVIDIA — Deep Learning Algorithm Engineer

    TensorRT-LLM · Deep Learning Inference · Santa Clara, CA

    SGLang DeepSeek-V3.2 NSA, Spec-V2/EAGLE, GB300 long-context, multi-node Pareto sweeps.

  2. May 2024 — Aug 2024

    NVIDIA — System Software Engineer Intern

    Autonomous Vehicle ML & Driveworks Platform · Santa Clara, CA

    RoadCast 2.1 — distributed data logging with dwProto serialization, batched I/O, NVSci/TCP/DFS.

  3. Feb 2023 — May 2023

    Amazon Web Services — Cloud Development Associate Intern

    Cloud Architecture & MLOps · Beijing

    TB-scale data lake (Kinesis → S3 → Glue/Athena), SageMaker MLOps with CodePipeline + EKS/ECS.

  4. Sep 2022 — Dec 2022

    ByteDance / TikTok — R&D Engineer Intern

    Distributed Video Quality DL Evaluation · Shanghai

    Multimodal DL/ML video QoE evaluation (D-VQA, multimodal Transformer), Spark on HDFS.

  5. 2023 — 2024

    Carnegie Mellon University

    M.S. Information Technology · Information Networking Institute · GPA 3.84/4.00

  6. 2019 — 2023

    Tongji University

    B.E. Computer Science & Technology · GPA 90.8/100 · First-class Scholarship

06 — Writing

Notes from the road

A few longer reflections on projects I’ve been part of — currently lives on LinkedIn.