Mengchi Zhang
Research Scientist at Meta
Menlo Park, California, United States
Summary
🤩
RockstarMengchi Zhang is a Research Scientist with a decade of experience in GPU-focused systems and ML engineering, based in Menlo Park. At Meta and across open-source projects, he has optimized GPU-accelerated routines and embedding workflows—contributing practical performance improvements to Facebook's FBGEMM and adding int8 dequantize kernels and batched embedding support. He has deep experience with GPU simulation and PTX-level tooling from work on GPGPU-Sim, where he improved CUDA dynamic parallelism handling, refactored parsers, and modernized compatibility for newer Python environments. Mengchi combines low-level systems debugging with benchmarking and testing rigor, ensuring changes are measurable and production-ready. Colleagues would note his knack for making subtle IR and parser fixes that unlock broader simulator and library capabilities rather than only flashy features.
10 years of coding experience
Github Skills (17)
pytorch10
c-language10
fba10
emm10
gpu-programming10
machine-learning10
f10
cuda10
cprogramming-language10
parser9
parsing9
parse9
quantization9
benchmarking8
build-system8
Programming languages (10)
Github contributions (5)
gpgpu-sim/gpgpu-sim_distribution
May 2017 - Sep 2019
GPGPU-Sim provides a detailed simulation model of contemporary NVIDIA GPUs running CUDA and/or OpenCL workloads. It includes support for features such as TensorCores and CUDA Dynamic Parallelism as well as a performance visualization tool, AerialVisoin, and an integrated energy model, GPUWattch.
Role in this project:
Back-end Developer
Back-end DeveloperContributions:143 commits, 6 PRs, 5 comments in 2 years 4 months
Contributions summary:Mengchi Zhang's contributions primarily focused on enhancing the GPGPU-Sim simulator's functionality related to PTX (Parallel Thread Execution) code and the simulation of GPU behavior. They addressed issues in the PTX intermediate representation (IR) and improved the handling of features like predicated returns and CUDA dynamic parallelism. The user implemented changes for better compatibility with the new Python version, and introduced options for caching. Furthermore, they refactored the cuobjdump parser and integrated it to enable more features.
simsimulationgpgpuopenclnvidia
pytorch/FBGEMM
Oct 2021 - Jan 2023
FB (Facebook) + GEMM (General Matrix-Matrix Multiplication) - https://code.fb.com/ml-applications/fbgemm/
Role in this project:ML Engineer
ML EngineerContributions:13 reviews, 15 commits, 46 PRs in 1 year 3 months
Contributions summary:Mengchi's contributions primarily revolve around modifying and adding functionalities related to the FBGEMM library, specifically targeting GPU-accelerated operations. They modified the `merge_pooled_embeddings` function, optimizing it for GPU usage and incorporated batch size as a direct parameter. The user also added an int8 dequantize kernel to support batched mixed dimension pooled embedding output. Furthermore, the user benchmarked and added tests to ensure functionality.
matrix-multiplicationfacebookmultiplicationmatrixml-applications
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.