Mark Saroufim

Career

Mark Saroufim's homepage presents three current roles: co-founder of Core Automation, PyTorch maintainer, and co-founder of GPU MODE, a community formerly known as CUDA MODE. Saroufim is listed with Core Automation on the keynote program for PyTorch Conference North America 2026, with a session titled "Linear Algebra for the Age of Research."¹⁴

Saroufim's homepage lists earlier, pre-PyTorch projects on high-dimensional geometry and game AI built with reinforcement learning, and Saroufim co-authored a 2020 paper on parallel training of deep networks with local updates. A 2015 arXiv paper by Saroufim studies spatial trees, high-dimensional reductions and batch nearest-neighbor search.¹⁵⁶

Open-source work

The homepage groups Saroufim's PyTorch projects together, including torchao, TorchServe, the TorchX Ray scheduler, the launch of PyTorch 2.0, the PyTorch fast series, Project Popcorn, and the NeurIPS LLM efficiency and Hacker Cup AI challenges. PyTorch's maintainers page lists Saroufim as an emeritus maintainer of quantization (torch/ao) and among the TorchServe maintainers.¹²

Saroufim is a co-author of the PyTorch 2 paper (ASPLOS 2024), which describes TorchDynamo and TorchInductor, and of the 2025 TorchAO paper, which presents quantization and sparsity tools spanning pre-training, fine-tuning and serving. The NeurIPS 2023 LLM efficiency challenge, listed on the homepage, asked entrants to fine-tune a foundation model on a single GPU within one day.³⁷¹⁸

Work and contributions

• PyTorch (quantization and TorchServe) Open-source code

  Listed on PyTorch's maintainers page as an emeritus quantization (torch/ao) maintainer and among the TorchServe maintainers; co-author, listed with Meta, of the PyTorch 2 paper introducing TorchDynamo and TorchInductor. The repository is under a BSD-style license.²³⁹

• torchao Open-source code

  Listed on Saroufim's homepage under "Work on PyTorch"; the repository describes torchao as PyTorch-native quantization for training and inference, under a BSD-style license.¹¹⁰

• TorchAO: PyTorch-Native Training-to-Serving Model Optimization Published research

  Co-author of the 2025 paper describing torchao as a PyTorch-native framework that uses quantization and sparsity across a training-to-serving workflow.⁷

Timeline

1. Jul 13, 2015Posted a single-author arXiv paper on spatial trees and batch nearest-neighbor search.⁶
2. Dec 7, 2020Co-authored "Parallel Training of Deep Networks with Local Updates" on arXiv.⁵
3. 2023The NeurIPS 2023 LLM efficiency challenge, listed among Saroufim's PyTorch work, ran with a one-GPU, one-day fine-tuning format.⁸¹
4. Apr 2024Co-authored the PyTorch 2 paper on TorchDynamo and TorchInductor, presented at ASPLOS 2024.³
5. Jul 21, 2025Co-authored the TorchAO paper on PyTorch-native model optimization.⁷
6. Aug 20, 2026Announced as a keynote speaker for PyTorch Conference North America 2026, representing Core Automation.⁴

Other documented roles

Roles the sources document that are not confirmed as current.

• Listed with Meta as an author of the PyTorch 2 paper (2024), Meta³

Where they work

• Core Automation 1 person in the wiki
• GPU MODE 1 person in the wiki

Related people

• Kaichao YouInference and serving · ML frameworks and libraries
• Johannes GäßlerInference and serving · ML frameworks and libraries
• Lin QiaoInference and serving · ML frameworks and libraries
• Gavin UbertiInference and serving · ML frameworks and libraries
• Tim DavisML frameworks and libraries · Inference and serving
• Awni HannunML frameworks and libraries · Inference and serving
• Tianqi ChenInference and serving · ML frameworks and libraries
• Jeff JohnsonInference and serving · ML frameworks and libraries
• Patrick von PlatenML frameworks and libraries · Inference and serving
• Ian BuckML frameworks and libraries · Inference and serving