Xiangrui Yu

Xiangrui Yu 余湘锐

Pretraining Infrastructure Engineer, Baidu · Beijing

Email Google Scholar GitHub

About

I am a pretraining infrastructure engineer at Baidu (BMU), where I own the optimizer infrastructure for ERNIE. I bring new optimizers, such as Muon-style orthogonalized updates, into large-scale distributed training and make them fast: sharded optimizer states, fused kernels, and checkpoint save, load and resharding. I also work on distributed training performance in general: parallel strategies, communication overlap, and GPU throughput.

Before Baidu, I received my MPhil from HKUST (Guangzhou), where I worked on efficient LLM inference and sparse computation on GPUs.

Experience

Baidu, BMU Pretraining Infrastructure Optimizer Infrastructure Owner for ERNIE
  • Built paddle.optimizer.Muon: Newton–Schulz orthogonalized momentum for 2D weights, AdamW for the rest (source).
  • Built MuonShardingOptimizer: sharding stage 1 for Muon, which assigns whole matrices to owner ranks so orthogonalization stays local and shards the other parameters element-wise.
  • Contributor to the ERNIE 5.0 pretraining.
2025 – Present

Education

The Hong Kong University of Science and Technology (Guangzhou) MPhil, Data Science and Analytics Thrust
2023 – 2025
China University of Petroleum, Beijing B.Eng. in Computer Science and Technology
2019 – 2023

Publications

Full list on Google Scholar.

ICS 2026
DynSpAttn: Efficient Attention via Dual-Side Dynamic Sparsity on Sparse Tensor Cores
X. Yu, R. Fan, Z. Li, W. Luo, G. Gong, X. Chu
40th ACM International Conference on Supercomputing (ICS), 2026
ASPLOS 2026
ZipServ: Fast and Memory-Efficient LLM Inference with Hardware-Aware Lossless Compression
R. Fan, X. Yu, X. Pan, Z. Li, W. Luo, Q. Wang, W. Wang, X. Chu
31st ACM International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS), 2026
arXiv Code
PPoPP 2026
ROME: Maximizing GPU Efficiency for All-Pairs Shortest Path via Taming Fine-Grained Irregularities
W. Luo, Y. Chen, X. Yu, Q. Wang, R. Fan, H. Liu, X. Chu
31st ACM SIGPLAN Annual Symposium on Principles and Practice of Parallel Programming (PPoPP), 2026
arXiv 2026
ERNIE 5.0 Technical Report
ERNIE Team (contributor)
arXiv:2602.04705, 2026
arXiv
EuroSys 2025
SpInfer: Leveraging Low-Level Sparsity for Efficient Large Language Model Inference on GPUs
R. Fan, X. Yu, P. Dong, Z. Li, G. Gong, Q. Wang, W. Wang, X. Chu
20th European Conference on Computer Systems (EuroSys), 2025
Best Paper Award
PDF Code
CCGrid 2023
Balancing Computation and Communication in Distributed Sparse Matrix-Vector Multiplication
H. Mi, X. Yu, X. Yu, S. Wu, W. Liu
23rd IEEE/ACM International Symposium on Cluster, Cloud and Internet Computing (CCGrid), 2023
Paper Code

Awards