About Me

Hello Friend, welcome to my blog!

Contact:

• Email: zhyanwentao@outlook.com
• Github: yewentao256
• LinkedIn: Wentao Ye

1. Patent

• One-iter Tool (CN117312173A), patented in 2023, reducing model accuracy validation time from hours to minutes.

2. Awards

• 100k Cornell Startup Award
  • PolyRook: Fast 3D environment generation.
• National Third Prize
  • China University Computer Contest, WeChat Big Data Challenge, 2021
  • Rank 80 / 6,768 teams
  • Certificate
• National Third Prize
  • China University Computer Contest, Huawei Big Data Challenge, 2020
  • Rank 27 / 1,491 teams
  • Certificate
• National Second Prize
  • China Service Outsourcing Innovation & Entrepreneurship Competition, 2020
  • Top 1% / 6417 teams
  • Certificate
• National First Prize
  • China University Computer Capability Challenge, 2019
  • Runner-up / 414 teams
  • Certificate

3. Experience

• Machine Learning Engineer

  • Red hat
  • Jun 2025 - Current

• Deep Learning Engineer

  • SenseTime | SenseCore
  • Jul 2022 - Aug 2024

• R&D Intern

  • SenseTime | Research Institute (Deep Learning Frameworks)
  • Jan 2021 - Jul 2022

• Co-founder & CTO

  • Wuhan Hongyuan Investment & Technology Services Co., Ltd.
  • Nov 2019 - Sep 2020

• Co-founder

  • Yuye (Wuhan) Technology Development Co., Ltd.
  • Jun 2019 - Nov 2019

4. Education

• Master of Computer Science

  • Cornell University | New York, USA
  • May 2025
  • GPA: 4.21/4.0 (4.3 for A+)

• Bachelor of Software Engineering

  • Wuhan University | Wuhan, China
  • Jun 2022
  • GPA: 3.91/4.0
  • National Scholarship (top 1%)

5. Selected OpenSource Projects

Contributor

PyTorch

• 
• May 2023 - Present
• Optimized the cuDNN Convolution and cuDNN BatchNorm operators, achieving a 15% performance boost in CNN training and inference for computer vision tasks
• 30+ contributions to Pytorch.
• Authored a blog series with 15+ articles**, providing the developer community with insights into PyTorch’s core architecture and optimizations.
• Details at My Contributions

Maintainer

vLLM

• 
• Jun 2025 - Current
• Code owner for quantization, batch-invariant execution, caching, weight loading and CUDA kernels
• Led design and implementation of batch-invariant, showcased in the vLLM blog and mentioned at PyTorch Conference 2025.
• Optimized MoE shared-expert overlap scheduling, improving end-to-end throughput by ~6% and reducing time-to-first-token latency by 25%+.
• Integrated and tuned DeepGEMM on B200/H100 GPUs, delivering ~11% throughput gains on B200 and ~6% on H100 while preserving accuracy; Shipped DeepSeek V3.2 support in one week.
• Developed and optimized low-precision quantization kernels (INT8/FP8) for LLM inference, speeding up models by ~13% on H100 and FP8 by ~7% on B200 without accuracy loss.
• Details at My Contributions and Bi-weekly Journal

LazyLLM

• 
• May 2024 - Aug 2024
• Built a Retrieval-Augmented Generation (RAG) system with a specialized tree architecture, which improved query performance by 50% over LlamaIndex by enhancing the efficiency of parent/child node retrieval.
• Details at My Contributions

MMCV & PAVI Logger

• 
• Jan 2021 - Dec 2022
• Rebuilt the PAVI data collection SDK, achieving a 10× improvement in data upload efficiency through optimized parallel processing, significantly reducing ingestion time and enhancing performance for large-scale datasets.
• Integrated the proprietary PAVI Logger system into the MMCV library, enabling efficient and customizable logging for deep learning workflows, with the core system remaining private.

DeepLink & DIOPI

• Apr 2023 - May 2024
• Optimized Llama 2-70B training on 1024 NPUs by integrating distributed training strategies (ZeRO, Tensor Parallel, Pipeline Parallel) and operator-level optimizations. Achieved a 700% increase in TGS (Tokens/GPU/Second) and significantly boosted LLM performance.
• Details at Deeplink and DIOPI

Owner

GAN-Paint

• Nov 2024 - Jan 2025
• Developed a lightweight GAN (generative adversarial network) for large-area image completion and cross-scene stitching, achieving realistic outputs on a single RTX 2070 GPU.
• Implemented an end-to-end training pipeline with efficient data preprocessing, masking strategies, and evaluation, completing model training within hours.

MicroTorch

• Jun 2023 - Aug 2024
• Developed a minimalistic deep learning framework inspired by PyTorch, implementing core functionalities such as AutoGrad, dynamic computation graphs, and tensor operations.
• Designed to be lightweight and modular, making it ideal for educational purposes, with extensive examples to facilitate learning.

CMU CSAPP

• Dec 2022 - Feb 2024
• Self-studied the CMU CSAPP-15213 course and completed its associated labs, covering core concepts such as assembly optimization, multi-level cache, compiling and linking, exception control flow, virtual memory, and system-level I/O.
• Blogs

TinyNN

• Nov 2022 - Dec 2022
• Built TinyNN, a minimal implementation of Fully Connected Neural Networks and Convolutional Neural Networks, designed for educational and experimental purposes.

You After Taking Drugs

• Aug 2021
• Independently developed this system in 7 days using computer vision algorithms; optimized for smooth performance on a single i3 CPU, ensuring a seamless user experience and earning client approval in the first review.
• Software Copyright: “After Taking Drugs (Facial Human Morphing Experience)” (2022SR0021854).

Sicpy Compiler

• Nov 2020 - Dec 2020
• Designed and implemented an untyped programming language Sicpy and its corresponding compiler using flex and bison.
• Developed features including lexical, syntax, and semantic analysis, as well as type inference and automatic garbage collection via reference counting, providing a complete custom language framework for functional and imperative programming experimentation.

New Super Mario

• Apr 2020
• Group project with Jifeng Wu, Jinran Tang and Taihe Li.