• Supervisor: Prof. Hailong Yang
• Research Interests: Machine Learning System, Heterogeneous Computing

• GPA: 3.83/4.0, Rank: 13/267

Platform of Artificial Intelligence, Alibaba Cloud, Hangzhou. (07/2024 - Now)

R&D Engineer

• Core contributor of Llumnix, a dynamic scheduling system for large language model serving. [Code]
• Contributor and core maintainer of EasyCkpt, a high performance asynchronous checkpointing framework for large language model training, accelerating Qwen2.5 and Qwen3 training. [Doc]

Platform of Artificial Intelligence, Alibaba Cloud, Beijing. (06/2023 - 06/2024)

Intern R&D Engineer

• Designed and developed the prototype system of Llumnix. Llumnix proposes a dynamic scheduling policy that reschedules requests and their in-memory states with an efficient and scalable live migration mechanism, improving load balancing and isolation, mitigating resource fragmentation, and differentiating request priorities and SLOs.
• Llumnix: Dynamic Scheduling for Large Language Model Serving (OSDI’24, co-first author) [Paper/Slides/Talk]

Model-ToolChain Team, Sensetime, Beijing. (03/2023 - 05/2023)

Intern System Researcher

• Contributed to the training and inference optimization of the stable diffusion models on SenseMirage platform.
• Reproduced the training activation memory compression optimization proposed in Gist (framework presented in ISCA’18, not open sourced).

PaddlePaddle Team, Baidu, Beijing. (06/2022 - 09/2022)

Intern R&D Engineer

• Participated in the CCF-GLCC Open Source Programming Summer Camp, optimized the max_pool3d and depthwise_conv2d_grad operators in the Paddle framework.
• Achieved performance improvements of 224% and 36% respectively, surpassing PyTorch performance by 14% and 48%, submitt a total of 4 PRs to both Paddle (3) and PyTorch (1).

Platform of Artificial Intelligence, Alibaba Cloud, Beijing. (07/2021 - 12/2021)

Intern System Researcher

• Contributed to EasyScale, an elastic training system that achieves consistent model accuracy under resource elasticity for both homogeneous and heterogeneous GPUs.
• EasyScale: Elastic Training with Consistent Accuracy and Improved Utilization on GPUs (SC’23, fourth author) [Paper][Talk]

Exploiting Input Tensor Dynamics in Activation Checkpointing for Efficient Training on GPU

Jianjin Liao, Mingzhen Li, Hailong Yang, Qingxiao Sun, Biao Sun, Jiwei Hao, Tianyu Feng, Fengwei Yu, Shengdong Chen, Ye Tao, Zicheng Zhang, Zhongzhi Luan, Depei Qian

2023 IEEE International Parallel and Distributed Processing Symposium (IPDPS’23)

[PDF]

EasyScale: Elastic Training with Consistent Accuracy and Improved Utilization on GPUs

Mingzhen Li, Wencong Xiao, Hailong Yang, Biao Sun, Hanyu Zhao, Shiru Ren, Zhongzhi Luan, Xianyan Jia, Yi Liu, Yong Li, Depei Qian, Wei Lin

International Conference for High Performance Computing, Networking, Storage, and Analysis 2023 (SC’23)

[Paper][Talk]

Adapting Combined Tiling to Stencil Optimizations on Sunway Processor

Biao Sun, Mingzhen Li, Hailong Yang, Jun Xu, Huaitao Zhang, Zhongzhi Luan, Depei Qian

CCF Transactions on High Performance Computing 2023 (THPC’23)

[PDF]