Max Ryabinin
VP Of Research & Development, Model Shaping at Together AI
Amsterdam, North Holland, Netherlands
Summary
🤩
Rockstar🎓
Top SchoolMax Ryabinin is a VP of Research & Development (Model Shaping) at Together AI with a decade of experience building efficient, scalable deep learning systems and tooling for customizing open models. He leads research and product efforts around post-training quality, speedups for foundation models, and a Fine-Tuning Platform while previously driving core training infrastructure as a Distinguished Research Scientist. Max is the creator and an active maintainer of Hivemind, an open-source decentralized deep learning library used to train models across thousands of volunteer machines and integrated with frameworks like PyTorch Lightning. His contributions span low-level systems work—optimizing inference, deterministic dropout, k-bit quantization support—and education, where he teaches efficient deep learning systems and authors course materials. A PhD candidate in Computer Science based in Amsterdam, he combines top-tier research publications with hands-on engineering that moves distributed ML from theory into production.
10 years of coding experience
6 years of employment as a software developer
Doctor of Philosophy - PhD, Computer Science, Doctor of Philosophy - PhD, Computer Science at Higher School of Economics
Russian, English
Github Skills (29)
asynchronous10
algorithm10
pytorch10
language-model10
python10
optimizers10
pre-trained-model10
async10
machine-learning10
distributed-systems10
mlops10
deep-learning10
optimisation10
cuda10
quantization10
Programming languages (5)
Github contributions (5)
mryab/efficient-dl-systems
Dec 2021 - Jan 2023
Efficient Deep Learning Systems course materials (HSE, YSDA)
Role in this project:
ML Engineer
ML EngineerContributions:30 reviews, 19 commits, 20 PRs in 1 year 1 month
Contributions summary:Max contributed to the course materials for efficient deep learning systems, adding seminar and homework content related to CUDA, benchmarking, distributed training, and data parallelism. They provided example code and exercises demonstrating CUDA memory management, distributed training using PyTorch's `torch.distributed` module, and setting up data parallel training. The commits demonstrate a focus on practical implementation and educational content related to optimizing deep learning workflows.
deep-learninghsecudadistributed-trainingefficient-deep-learning
learning-at-home/hivemind
Mar 2020 - Jan 2023
Decentralized deep learning in PyTorch. Built to train models on thousands of volunteers across the world.
Role in this project:Back-end Developer & ML Engineer
Back-end Developer & ML EngineerContributions:20 releases, 487 reviews, 125 commits in 2 years 10 months
Contributions summary:Max made several changes to the core server and expert backend components of the project. Their work involved refactoring and modifying the `hivemind/server` and `hivemind/runtime` modules, with a focus on functionality and stability. Furthermore, they implemented deterministic dropout layers, showcasing expertise in enhancing model consistency and control over randomness, and adapted experts to support custom inputs and outputs. The contributions span across multiple areas of a distributed deep learning system.
pytorchhiveminddhtasynciovolunteers
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.