CV
Research, publications, systems projects, and selected experience.
Contact Information
| Name | Huanyi Xie |
| Professional Title | MS/PhD Student in Computer Science |
| huanyi.xie@kaust.edu.sa | |
| Phone | +966 54 941 9078 |
| Website | https://xiehuanyi.github.io |
Professional Summary
MS/PhD student at King Abdullah University of Science and Technology (KAUST), advised by Prof. Di Wang. I work on efficient training and inference of large language models, including LLM serving, memory-efficient fine-tuning, and data attribution.
Education
Publications
-
2026 Efficient Text-Attributed Graph Learning through Selective Annotation and Graph Alignment
TMLR
Huanyi Xie, Lu Yu, Tianhao Huang, Longfei Li, Meng Li, Jun Zhou, Lijie Hu, Di Wang.
-
2026 High-Throughput and Memory-Efficient Zeroth-Order Fine-tuning LLMs with Distributed Parallel Computing
ACL
Liangyu Wang, Huanyi Xie, Di Wang.
-
2026 Predicting LLM Output Length via Entropy-Guided Representations
ICLR
Huanyi Xie, Yubin Chen, Liangyu Wang, Lijie Hu, Di Wang.
-
2026 Dissecting Representation Misalignment in Contrastive Learning via Influence Function
ICLR
Huanyi Xie, Chenyang Ren, Khouloud Saadi, Shu Yang, Zhen Tan, Jingfeng Zhang, Lijie Hu, Di Wang.
-
2026 Evaluating Data Influence in Meta Learning
ICLR
Chenyang Ren, Huanyi Xie, Shu Yang, Meng Ding, Dongrui Liu, Lijie Hu, Di Wang.
-
2025 ZO2: Scalable Zeroth-Order Fine-Tuning for Extremely Large LMs with Limited GPU Memory
COLM
Liangyu Wang, Jie Ren, Hang Xu, Junxiao Wang, Huanyi Xie, David E. Keyes, Di Wang.
-
2025 Semi-supervised Concept Bottleneck Models
ICCV
Lijie Hu, Tianhao Huang, Huanyi Xie, Xilin Gong, Chenyang Ren, Zhengyu Hu, Lu Yu, Ping Ma, Di Wang.
Open Source Projects
-
nanoSGLang
Minimal LLM Serving Engine.
- Built a high-performance LLM serving engine from scratch in Python/PyTorch, implementing continuous batching, chunked prefill, paged KV cache, prefix-sharing radix cache, streaming API, and request-level metrics.
- Integrated FlashInfer paged attention and fused operators; benchmarked Qwen2.5-0.5B at 3469 tok/s on RTX A5000, exceeding the referenced SGLang baseline in the same workload.
-
nanoMegatron
Minimal Distributed Training Framework.
- Implemented core distributed-training algorithms in a compact educational codebase, including DDP, ZeRO-1/2/3, tensor parallelism, sequence parallelism, pipeline parallelism, and expert parallelism.
- Built a working SFT pipeline for Phi-tiny-MoE and GSM8K, with MoE AllToAll dispatch, ZeRO/FSDP-style sharding, and benchmark logs against DeepSpeed/PyTorch FSDP.
Awards
-
2023 Honorable Mention, Mathematical Contest in Modeling
MCM
-
2022 Baidu PaddlePaddle Developer Groups Gold Team Leader
Baidu
-
2022 Third Prize, Baidu Cognitive AI Creative Contest
Baidu
-
2021 Second Prize, National College Student Mathematics Competition (Non-Mathematics Category)
Chinese Mathematical Society
-
2021 First Prize, China Collegiate Computing Contest Artificial Intelligence Innovation Contest
CCCC