Peizhen Zheng

AI Infrastructure Researcher

Huawei Technologies · 2012 Laboratories

About

I am an AI Infrastructure Researcher in the Computational Mathematics Group, Theory Research Department, 2012 Laboratories at Huawei Technologies. My work spans LLM training, inference and data infrastructure.

I independently develop APIE, an LLM inference framework designed for recursive self-improvement (RSI), with optimization skills, benchmarks and validation tools for coding agents. I have also worked on openPangu MTP architecture selection and full-graph inference, DeepSeek performance optimization on Ascend, and distributed pretraining data pipelines. My engineering interests cover GPU/NPU systems, serving, communication overlap and data processing.

I received my master’s degree from Georgia Tech and studied at Zhejiang University’s ZJU-UIUC Institute. My earlier research interests include computer graphics and 3D scene reconstruction.

Download resume · Google Scholar · GitHub

Interests

LLM inference & distributed computing
RSI & agent-driven optimization
Pretraining data infrastructure
Computer graphics & 3D reconstruction

Education

MS in Computer Science and Engineering, 2024
Georgia Institute of Technology
BS in Civil Engineering, 2022
Zhejiang University · ZJU-UIUC Institute

Experience

AI Infrastructure Researcher

Huawei Technologies · 2012 Laboratories

February 2025 – Present Shanghai

Computational Mathematics Group, Theory Research Department (Yuwei Fan).

Work on inference engines, MTP, communication optimization and pretraining data infrastructure.

Selected projects

Earlier graphics work and all projects

openPangu-v3: agent-assisted distributed pretraining data pipelines

Independently build arXiv parsing and Data-Juicer-based GitHub processing, combining distributed deduplication with human–agent error-driven iteration.

APIE / RSI Infra: LLM inference for recursive self-improvement

An independently developed inference framework with optimization skills, benchmarks, fused graphs, asynchronous speculation and Rust parallel tokenization.

openPangu-v2: pretraining architecture selection and MTP inference optimization

Core contributions to three-head MTP selection, 2048-token sliding windows and full-graph speculative execution for long contexts.

DeepSeek-V3: inference performance on Ascend

Core work on lossless parallel tokenization and FlashComm communication/stream adaptation, with LoPT published at ACL 2026.

Publications & research

Google Scholar: academic profile

LoPT: Lossless Parallel Tokenization Acceleration for Long Context Inference of Large Language Model

ACL 2026 main conference long paper. Exact-output parallel tokenization for long-context inference through character-position matching and adaptive chunk expansion.

Wei Shao, Zheng Lingchao, Pengyu Wang, Peizhen Zheng, Li Jun, Yuwei Fan

3D Gaussian Splatting against Moving Objects for High-Fidelity Street Scene Reconstruction

Street-scene reconstruction with transparency control and progressive Gaussian refinement. Archived research; the arXiv record is withdrawn.

Peizhen Zheng, Dongjing Jiang, Qingchong Jiao, Redouane EL Bouchtaoui, Flynnwell Jianfei Zhang

Skills

Platforms — GPU / NPU; PyTorch, torch_npu and Triton; Ascend adaptation and performance optimization.

Inference — Continuous batching, paged KV cache, P/D disaggregation, MTP, fused graphs and asynchronous pipelines; vllm-ascend (ascend-vllm/omni-infer).

Training & parallelism — Familiar with Megatron and PyTorch FSDP; DP / TP / PP / CP / SP / EP; collectives and computation–communication overlap.

Development & data — Python, C++, Rust and Git; Data-Juicer, distributed pipelines, document parsing and code deduplication.

Agent development — Claude Code / Codex / DeepSeek Harness; skills, issue/PR workflows and automated validation.

Languages — TOEFL 107; CET-6.