Omkar Salpekar
Member Of Technical Staff at OpenAI
San Francisco Bay Area United States
Summary
🤩
Rockstar🎓
Top SchoolOmkar Salpekar is a Member of Technical Staff focused on research and pretraining of large language models, currently at OpenAI after leading pretraining and RL scaling efforts at Meta. Over a decade he’s built distributed training and MLOps infrastructure—scaling Llama pretraining to 100k+ GPUs, founding a 20-engineer team for fault-tolerant massive-scale training, and delivering production reliability for PyTorch. His open-source contributions span PyTorch, torchvision, and Modin, where he improved CI/CD, Windows/Python packaging, and core dataframe operations for large-scale data workflows. Equally comfortable in research and production, he has published systems work (VLDB, MLSys) and driven multi-million-dollar cost savings via efficient fault tolerance and elasticity. Early career research at Berkeley RISE Lab and hands-on roles from compiler work to distributed systems give him a rare blend of ML research depth and production-grade engineering. He also co-founded and ran campus accelerator and nonprofit programs, demonstrating a persistent drive to scale impact beyond code.
10 years of coding experience
13 years of employment as a software developer
Master of Science - MS Machine Learning Data Science, Master of Science - MS Machine Learning Data Science at University of California, BerkeleyMoreau Catholic High School
Stackoverflow
1reputation
0reached
0answers
0questions
Github Skills (33)
pytorch10
github-ci10
c-language10
docker10
python10
data-science10
testing10
tensorrt10
dataframes10
pandas10
machine-learning10
circleci10
dockers10
cicd10
dataframe10
Programming languages (6)
Github contributions (5)
pytorch/pytorch
Jan 2020 - Jan 2023
Tensors and Dynamic neural networks in Python with strong GPU acceleration
Role in this project:
ML Engineer
ML EngineerContributions:182 reviews, 130 commits, 172 PRs in 3 years
Contributions summary:Omkar primarily contributed to the optimization and maintenance of PyTorch, focusing on areas related to machine learning model execution and performance. Their work involved supporting new operations within the `fx2trt` module for TensorRT integration, which accelerates deep learning models. They also addressed compiler errors and fixed build issues, showcasing their expertise in debugging and maintaining the codebase. Additionally, they modified code related to the removal of unused variables within the SDPA project.
pythongpu-accelerationdeep-learninggpunumpy
facebookincubator/gloo
Dec 2020 - Feb 2021
Collective communications library with various primitives for multi-machine training.
Role in this project:Backend Engineer
Backend EngineerContributions:3 reviews, 6 commits, 11 PRs in 2 months
Contributions summary:Omkar contributed to the Gloo project by implementing and improving core functionality. They worked on computing local ranks for multi-GPU training, adding copyright headers to multiple files, and renaming variables/adding comments in multiproc tests for clarity. Furthermore, the user refactored the codebase by making "localhost" the default device, and validating context timeout inputs.
communicationsmachine-learningtrainingprimitivescollective
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.