Scholar
Pengfei Zheng
Google Scholar ID: 3QxdztoAAAAJ
Huawei Technologies
Citations & Impact
Citations
241
H-index
5
i10-index
4
Publications
12
Co-authors
12
list available
Contact
Publications
Resume (English only)
Academic Achievements
- arXiv preprint: 'Speculative MoE: Communication Efficient Parallel MoE Inference with Speculative Token and Expert Pre-scheduling'
- 'CoffeeBoost: Gradient Boosting Native Conformal Inference for Bayesian Optimization' accepted to AAAI 2025
- 'Centrum: Model-based Database Auto-tuning with Minimal Distributional Assumptions' accepted to SIGMOD 2025
- 'Mirage: MOE + Decision Transformer for non-interruption, non-overlap resource provision in ML training' published at SC 2023
- 'Shockwave: Fair and Efficient Scheduling for Dynamic Adaptation in Machine Learning' published at NSDI 2023
- Mentored students including Lynn Liu (HyperAPX, now Ph.D. at UC Berkeley), Rui Pan (Shockwave, now Ph.D. at Princeton), and Calvin Ma (Hound)
Background
- Full-time Staff Researcher at Huawei Technologies
- Research areas include autonomous system agents, distributed systems, and disaggregated datacenter architecture
- Designs statistical and neural learning methods to model dynamics and uncertainty in large-scale distributed computer systems
- Develops algorithmic decision-making mechanisms (e.g., convex/nonlinear optimization, Bayesian optimization, contextual bandits, reinforcement learning) to optimize system performance, efficiency, and scalability
- Recent work includes black-box optimization for autonomous system tuning, real-time recommendation algorithms for preconditioner/solver selection in linear systems, dynamic market theory, and stochastic policy learning for scheduling and resource allocation
- Works on high-performance ML training and inference, building multi-dimensional hybrid parallelism solvers (data, tensor, pipeline, sequence, expert parallelism, offload, rematerialization, pipeline interleaving) to optimize token throughput, MFU, and scale-out linearity for hyper-scale LLM training on massive GPU/NPU clusters
- Researches dynamic MoE layers and real-time model/input pruning techniques to accelerate LLM inference
