Ranran Haoran Zhang

PhD Candidate, Penn State University

profile_new.jpg

I am a Ph.D. candidate in Computer Science and Engineering at Penn State University, advised by Dr. Rui Zhang. My research focuses on LLM inference and AI systems, connecting research and engineering to make models more useful, dependable, and accessible in practice. My current work spans:

  • Inference systems and runtimes. As a core maintainer of vLLM-metal, I develop efficient serving on Apple Silicon, working across attention kernels, scheduling, memory management, and multi-machine execution.
  • Speculative decoding and draft-model training. I investigate decoding correctness and contribute to Speculators, improving training’s numerical accuracy, memory efficiency, and speed.
  • Evaluation and benchmarking. I developed SiliconBench to evaluate inference engines on serving speed, memory use, and output fidelity under practical workload and hardware constraints.

My broader background includes learning with imperfect annotations, transfer across NLP tasks, and production AI deployment at eBay. Across these areas, I follow meaningful problems into the parts of the stack where they can be solved.

Previously, I obtained my M.S. in Information Management from the University of Illinois Urbana-Champaign, advised by Dr. Heng Ji, and my B.S. in Computer Science from Changsha University of Science & Technology, advised by Dr. Daojian Zeng.

news

Sep 12, 2026 New preprint: SiliconBench, our benchmark of LLM serving speed, memory use, and output fidelity on unified-memory desktops.
Jun 12, 2026 New blog post: One Launch for Any Batch: The Binary Search Inside vllm-metal’s Varlen Attention.

selected publications

  1. arXiv
    SiliconBench: Speed, Memory, and Fidelity for LLM Serving on Unified-Memory Desktops
    Ranran Haoran Zhang, Aysa X. Fan, David Munhá Correia, and 2 more authors
    arXiv:2609.19169, 2026
    Under review at NeurIPS 2026 Evaluations & Datasets Track.
    arXiv Bib HTML Website
  2. EMNLP
    Correctness Forensics for Batch Speculative Decoding: Diagnosing the Ragged Tensor Problem
    Ranran Haoran Zhang, Soumik Dey, Ashirbad Mishra, and 3 more authors
    In Findings of the Association for Computational Linguistics: EMNLP 2026, 2026
    Accepted.
    arXiv Bib HTML
  3. NAACL
    From Lazy to Prolific: Tackling Missing Labels in Open Vocabulary Extreme Classification by Positive-Unlabeled Sequence Learning
    Ranran Haoran Zhang, Bensu Uçar, Soumik Dey, and 3 more authors
    In Findings of the Association for Computational Linguistics: NAACL 2025, 2025
    arXiv Bib HTML
  4. EMNLP
    Minimize Exposure Bias of Seq2Seq Models in Joint Entity and Relation Extraction
    Ranran Haoran Zhang*, Qianying Liu*, Aysa X. Fan, and 5 more authors
    In Findings of the Association for Computational Linguistics: EMNLP 2020, 2020
    Bib HTML
  5. AAAI
    CopyMTL: Copy Mechanism for Joint Extraction of Entities and Relations with Multi-Task Learning
    Daojian Zeng*, Ranran Haoran Zhang*, and Qianying Liu
    In Proceedings of the AAAI Conference on Artificial Intelligence, 2020
    arXiv Bib HTML

View all publications