Taolue Kurt Yang ski touring in the mountains

Taolue (Kurt) Yang

Ph.D. Candidate in Computer Science, Temple University

I build systems that make AI and data-intensive science faster and cheaper by removing redundancy: compressing data, lowering numerical precision, and avoiding repeated computation, while keeping results exact or within guaranteed error bounds. I work mostly on GPUs, but also on TPUs and CPUs.

I am advised by Prof. Sian Jin and work closely with Prof. Xinghua (Mindy) Shi on genomic data. Along the way I have interned twice at Google and once at Oracle, and my changes to NVIDIA's NeMo-RL and ModelOpt were merged upstream. Outside of work, you will usually find me skiing.

Email CV Google Scholar GitHub LinkedIn

News

Research

Compression-native genomic computing Sequencing and pangenome data grow faster than the hardware to analyze them. I build GPU compressors for this data (CuVert-Q, GFAz, GPUFASTQLZ, SeqBench), run analytics directly on compressed pangenome graphs, and, most recently, share exact computation across thousands of personal genomes when running deep-learning models, making inference up to 584× faster than standard batched execution with identical outputs.
Quantization for efficient LLM training and inference I have worked on reduced-precision LLM systems across hardware: int8 inference kernels on AMD CPUs at Oracle, LLM serving on TPUs and GPU quantization at Google, and error-bounded compression of KV caches and embeddings in my Ph.D. work. Next, I am interested in quantization-aware RL post-training, FP8/FP4 fine-tuning, and error-bounded KV-cache quantization.

Publications

Full list on Google Scholar.

Conference Papers

  1. CuVert-Q: Resolving the Throughput-Ratio Tradeoff in Genomic Sequence Data Compression via GPU Acceleration T. Yang, Y. Liu, B. Jiang, C. Li, X. Shi, S. Jin ACM-BCB 2026Best Paper AwardPaperCode
  2. GFAz: State-of-the-Art Graphical Fragment Assembly Compression T. Yang, Y. Liu, B. Jiang, X. Shi, S. Jin ICS 2026PaperCode
  3. SeqBench: A Benchmark Suite for Lossless and Lossy Compression of Sequence Data T. Yang, Y. Liu, C. Li, X. Shi, S. Jin ACM-BCB 2024PaperCode
  4. TOPIQ: Statistical Error Propagation for Quantity-of-Interest Prediction under Lossy Compression Y. Liu, B. Jiang, T. Yang, S. Di, R. Underwood, S. Jin SC 2026
  5. 3D Gaussian Splatting for Scientific Particle Data Compression and Rendering B. Jiang, Y. Liu, T. Yang, S. Di, S. Jin SC 2026
  6. Error-Bounded Point Cloud Compression Using Truncated Octahedron Quantization Y. Liu, L. Zhang, R. Li, B. Jiang, T. Yang, K. Zhao, S. Di, E. Dragut, S. Jin VLDB 2026PaperarXivCode
  7. PackKV: Reducing KV Cache Memory Footprint through LLM-Aware Lossy Compression B. Jiang, T. Yang, Y. Liu, X. He, S. Di, S. Jin IPDPS 2026
  8. EmbdC: Error-Bounded Lossy Video Embedding Compression for On-Device LLM Inference B. Jiang, T. Yang, Y. Liu, X. He, S. Di, S. Jin CCGrid 2026
  9. Advancing Scientific Data Compression via Cross-Field Prediction Y. Liu, W. Jia, T. Yang, B. Jiang, M. Yin, S. Jin HPDC 2025

Workshop and Short Papers

  1. GPUFast-Q: Novel High-Throughput Sequence Data Compression on GPUs Using Fine-Grained Parallelism T. Yang, Y. Liu, B. Jiang, C. Li, X. Shi, S. Jin DCC 2026
  2. GPUFASTQLZ: An Ultra Fast Compression Methodology for Fastq Sequence Data on GPUs T. Yang, Y. Liu, B. Jiang, S. Jin SC24 Workshops
  3. Accurate Performance Modeling and Uncertainty Analysis of Lossy Compression in Scientific Applications Y. Liu, T. Yang, S. Jin DCC 2025Paper
  4. Enhancing Lossy Compression through Cross-Field Information for Scientific Applications Y. Liu, W. Jia, T. Yang, M. Yin, S. Jin SC24 Workshops

Under Review and In Preparation

  1. Computing over Compression: Pangenome Analytics Directly on Grammar-Compressed Graphs T. Yang, Y. Liu, B. Jiang, X. Shi, S. Jin Submitted to FAST 2027
  2. Exact Per-Layer Computation Sharing for Convolutional Sequence Models Across Haplotypes T. Yang, S. Jin In preparation
  3. KVComp: A High-Performance, LLM-Aware, Lossy Compression Framework for KV Cache B. Jiang, T. Yang, Y. Liu, C. Zhang, X. He, S. Jin PreprintarXiv

Industry Experience

Google · Cloud AI (Vertex AI)May – Aug 2026 · Sunnyvale, CA
Software Engineer Intern
  • Built a large-scale post-training quantization evaluation framework for LLMs across GPU generations, and used it to set hardware-aware deployment recipes.
  • Made quantization-aware RL training 70% faster in NVIDIA NeMo-RL by caching fake-quantized weights, and fixed a Megatron weight-folding bug in NVIDIA ModelOpt; both were merged upstream.
Google · AI and Infrastructure (MaxText)Jul – Oct 2025 · Kirkland, WA
Software Engineer Intern
  • Optimized Mixture-of-Experts LLM inference on TPUs, implemented Mixture of Block Attention in JAX, and built an OpenAI-compatible serving API for MaxText.
Oracle · MySQL HeatWaveJan – Jul 2025 · Redwood City, CA
Software Engineer Intern
  • Built an 8-bit quantized LLM inference stack for AMD CPUs with AVX-512 and int8 GEMM kernels, with 4.5× faster prefill than llama.cpp.
  • Developed a relaxed speculative decoding algorithm that speeds up generation by 1.5× (patent filed).

Open Source

Upstream contributions to NVIDIA NeMo-RL (#3226, #3441, #3556) and NVIDIA ModelOpt (#2131). My research software includes CuVert-Q, GFAz, and SeqBench.

Teaching

Lab instructor at Temple University for CIS 1068: Program Design and Abstraction, the first required programming course in the CS major (Spring 2024, two sections; Fall 2024), and for CIS 2109: Database Management Systems (Fall 2024). I ran weekly labs, taught Java, searching and sorting, and SQL, graded, and held office hours.

Service

Reviewer for ACDSA 2026 and 2027 and ICECET 2026.