Haozheng Fan

Bellevue, Washington, United States

Join Prog.AI to see contacts

Summary

🤩

Rockstar

🎓

Top School

Haozheng Fan is a Senior Applied Scientist with nine years of experience building high-performance ML infrastructure and deep learning compilers, currently working on MoE inference and LLM pre-training on AWS Trainium. He previously led design and implementation of RAF, traced PyTorch models with Project Razor, and delivered GPU speedups through CUTLASS integration for Amazon workloads. An active open-source contributor and committer for Apache TVM and MXNet, he has advanced compiler tooling—adding i64 support, CUDA reduction scheduling, and BroadNumpy-style operator coverage that enabled 200+ NumPy ops on GPU. Haozheng combines low-level systems optimization with large-scale model training expertise, and his work underpins recent Trainium-backed papers and production search model improvements. Notably, he reduced FFI invocation overhead by 5x in MXNet through a pragmatic C++/Python interface, showing both research depth and production-minded engineering.

9 years of coding experience

5 years of employment as a software developer

Bachelor of Engineering - BE Computer Science, Bachelor of Engineering - BE Computer Science at Zhejiang University

Github Skills (23)

algorithm10

algorithms10

c-language10

python10

mxnet10

machine-learning10

deep-q-learning10

relayer10

deeplearning-ai10

compiler-compiler10

deep-learning10

cuda10

compiler10

cprogramming-language10

relay10

Programming languages (3)

C++CPython

Github contributions (5)

apache/mxnet

Aug 2019 - May 2020

Lightweight, Portable, Flexible Distributed/Mobile Deep Learning with Dynamic, Mutation-aware Dataflow Dep Scheduler; for Python, R, Julia, Scala, Go, Javascript and more

Role in this project:

Back-end Developer & ML Engineer

Contributions:19 commits, 36 PRs, 3 pushes in 9 months

Contributions summary:Haozheng primarily contributed to the implementation and enhancement of TVM (Tensor Virtual Machine) integration within the MXNet framework. Their work involved modifying the source code to incorporate and improve TVM-related functionalities, including broadcast backward operations and infrastructure for operator attributes. The contributions include modifications to ufunc.cc and basic ufunc.py files and test scripts, suggesting improvements to low-level numerical computation and the incorporation of TVM features to the core of MXNet. This indicates a focus on optimizing the performance and expanding the capabilities of the deep learning framework.

pythonschedulerdataflowmutationdata-science

apache/tvm

Sep 2019 - Dec 2021

Open deep learning compiler stack for cpu, gpu and specialized accelerators

Role in this project:

Back-end Developer

Contributions:12 reviews, 9 commits, 14 PRs in 2 years 2 months

Contributions summary:Haozheng's contributions center around enhancing the TVM compiler stack, particularly in areas like Relay and TOPI. They focused on supporting i64 indices in Relay and implemented CUDA reduction scheduling within TOPI. Furthermore, they addressed and resolved issues related to bounds and casts within the arithmetic simplification and autodiff modules. The user's work improves the compiler's capabilities and performance.

metalvulkancompilertensoropencl

Find and Hire Top DevelopersWe’ve analyzed the programming source code of over 60 million software developers on GitHub and scored them by 50,000 skills. Sign-up on Prog,AI to search for software developers.

Request Free Trial