CV

Education

M.S. in Computer Science and Technology, Chongqing University, Sep. 2018 – Jun. 2021
- Thesis: Automatic Summarization System for Multi-domain Text
- Advisor: Prof. Jiang Zhong
B.S. in Computer Science and Technology, Chongqing University, Sep. 2014 – Jun. 2018
B.S. in Law (second degree), Chongqing University, 2018

Working experience

Jul. 2026 – present AI Framework Researcher
- Beijing Academy of Artificial Intelligence (BAAI) Beijing, China
- Leading Torch-FL, a PyTorch device plugin for the FlagOS stack: a unified flagos device with per-operator routing across vendor kernels, FlagGems/Triton compiler kernels, compatibility boxing, and CPU fallback.
Jul. 2021 – Jul. 2026 Senior AI Infrastructure Engineer / Project Lead
- Huawei, MindSpore Team Beijing, China
- MSAdapter / MindTorch (Project lead, Dec. 2024 – Jul. 2026)
  - Designed the compatibility stack for migrating PyTorch workloads to MindSpore + Ascend.
  - Enabled Pangu with no code changes and Day-0 ports of DeepSeek / Qwen.
  - Dynamic-graph out-of-the-box performance reaches about 95% of torch-npu, and exceeds it by about 10% after tuning.
  - Combined MSAdapter with PIJIT so MoE models such as DeepSeek-V3 can run high-performance inference without graph splitting.
- MindNLP (Project lead, Jul. 2022 – Jul. 2026)
  - Led architecture for pretrained model support and Hugging Face-style pipelines on MindSpore.
  - Optimized DeepSeek-OCR by consolidating expert computational graphs, raising NPU utilization from 8% to 30%.
  - Built an Apache Arrow + mmap preprocessing pipeline with about 4× throughput versus the native PyTorch DataLoader.
  - Supported official Hugging Face Transformers model classes on Ascend 910A/910B/310B, GPU, and CPU.
- MindSpore core engine (Core engineer, Jul. 2021 – Jul. 2026)
  - Designed Automatic Mixed Precision (gradient scaling and overflow detection) for stable Float16 training on Ascend 910.
  - Worked on a hybrid functional / OOP programming paradigm with JIT optimization comparable to JAX.
  - Implemented MindSpore RNN operators and standardized Transformer Encoder / Decoder / MHA interfaces.

Selected projects

Torch-FL Jul. 2026 – present
- PyTorch device plugin for FlagOS at BAAI. Users program against standard PyTorch APIs and device="flagos"; the plugin isolates vendor differences behind one logical device.
- Per-operator backends: native vendor kernels (e.g. ACLNN), CUDA/compatibility boxing, portable FlagGems/Triton kernels, and explicit CPU fallback.
- Covers eager execution, autograd, torch.compile, distributed collectives, and profiler integration, with support spanning CUDA, Ascend, MetaX, and other accelerators.
PocketLLM May 2026 – present
- Independent C++/CUDA and PyTorch inference stack for consumer multi-GPU systems. Started as DeepSeek-V4-Flash on 4× RTX 2080 Ti; now also covers MiniMax-M2.7, GLM-5.2, and Qwen3.8-27B-FP8.
- Custom kernels, low-bit formats, tensor/expert parallelism, and CPU-resident routed-expert staging on GPUs without native FP4/FP8 tensor cores.
- On 4× RTX 2080 Ti, C++ FP4 measures about 401 tok/s prefill and 3.7 tok/s decode with a 64k-token context.
- Docs · Technical report (PDF)

Research experience

CAAI–MindSpore Academic Research Funding Jun. 2024 – Jul. 2025
- Project: MoE inference optimization. Role: expert committee member and industrial mentor.
- Authored the official application guide; co-designed expert-aware prefetch and expert offload in Klotski; guided unified-memory deployment and prefetcher launch-cost optimization in Fate.
National Key Research and Development Program of China Nov. 2017 – Dec. 2020
- Project: key technologies for professional knowledge services. Role: machine learning algorithm engineer.
- Led NER, relation extraction, and automatic summarization; published DSMER at ECIR 2021.
- Built a labeling system and an internal NER dataset; about 80% NER accuracy and ROUGE 30+ on in-house summarization data.

Publications

Fate: Fast Edge Inference of Mixture-of-Experts Models via Cross-Layer Gate

Fang, Zhiyuan, Xingfan Yu, Yuegui Huang, Zicong Hong, Yufeng Lyu, Wuhui Chen, Yue Yu, and Fan Yu. "Fate: Fast Edge Inference of Mixture-of-Experts Models via Cross-Layer Gate." In Proceedings of the ACM Web Conference 2026, pp. 5537–5546.

Klotski: Efficient Mixture-of-Expert Inference via Expert-Aware Multi-Batch Pipeline

Fang, Zhiyuan, Yuegui Huang, Zicong Hong, Yufeng Lyu, Wuhui Chen, Yue Yu, Fan Yu, and Zibin Zheng. "Klotski: Efficient Mixture-of-Expert Inference via Expert-Aware Multi-Batch Pipeline." In Proceedings of the 30th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2, pp. 574–588. 2025.

ICDT: Incremental Context Guided Deliberation Transformer for Image Captioning

Lai, Xinyi, Yufeng Lyu, Jiang Zhong, Chen Wang, Qizhu Dai, and Gang Li. "ICDT: Incremental Context Guided Deliberation Transformer for Image Captioning." In Pacific Rim International Conference on Artificial Intelligence, pp. 444–458. 2022.

DSMER: A Deep Semantic Matching Based Framework for Named Entity Recognition

Lyu, Yufeng, and Jiang Zhong. "DSMER: A Deep Semantic Matching Based Framework for Named Entity Recognition." In European Conference on Information Retrieval, pp. 419–432. 2021.

Honors & awards

First-Class Scholarship, Chongqing University 2018 – 2020
Excellent TA Award, Chongqing University Mar. 2019
Third Prize, NSCSCC, Chinese Ministry of Education Sep. 2018
Huawei Scholarship Dec. 2017

Skills

Deep learning frameworks: internals of PyTorch and MindSpore; custom operators and high-performance model implementation.
Large-scale LLM training: Megatron-LM and DeepSpeed; pretraining and fine-tuning on clusters from 100 to 1,000+ GPUs/NPUs.
High-performance inference: vLLM and SGLang; framework-level changes and kernel-level optimization.

Yufeng Lyu

CV

Education

Working experience

Selected projects

Research experience

Publications

Fate: Fast Edge Inference of Mixture-of-Experts Models via Cross-Layer Gate

Klotski: Efficient Mixture-of-Expert Inference via Expert-Aware Multi-Batch Pipeline

ICDT: Incremental Context Guided Deliberation Transformer for Image Captioning

DSMER: A Deep Semantic Matching Based Framework for Named Entity Recognition

Teaching

Invited Guest Lecturer, LLM Training and Deployment Optimization with MindSpore

Guest Lecturer, LLM Development and Innovation Program

Online Teaching of Computer Organization

Teaching Assistant, Computer Organization and Integrated Hardware Design

Talks

Torch-FL: A Virtual Device Layer for PyTorch

Torch-FL: Ending the Fragmentation Nightmare of Heterogeneous AI Chips

LLM systems mini-lectures

An introduction to nested named entity recognition

CS graduate admissions, overseas applications, and internships

Life as a computer science graduate student

How to read an NLP paper

How to start a graduation project

From digital logic to computer organization

Honors & awards

Skills