Abstract:

In this keynote session from the PyTorch Conference, Matt White welcomes the audience back for the afternoon sessions before introducing Joe Spisak, Product Director at Meta. Spisak unveils Meta's next-generation, PyTorch-native AI development stack, designed from the ground up to address the challenges of scale and hardware heterogeneity. He outlines the three core principles behind the stack: being PyTorch-native and interoperable, building for massive scale (from training to edge inference), and supporting a diverse hardware landscape including custom silicon. Spisak introduces several new open-source projects that form this stack: Executor 1.0 for on-device AI, Torch Forge for reinforcement learning, Monarch for distributed orchestration, Helium for simplified high-performance kernel authoring, and TorchComms for unified communications across different hardware. He concludes with an invitation for the community to collaborate on these early-stage projects.

Meta's Next-Generation AI Stack for PyTorch

• 1:34 Introduction: Joe Spisak, Product Director at Meta, is introduced to make special PyTorch community announcements.
• 2:19 Core Principles for New AI Stack: Spisak outlines the three guiding principles for Meta's next-generation AI tools:
  • PyTorch Native: Tools must be interoperable and have the look and feel of PyTorch.
  • Built for Scale: The stack must support Meta's scale, from training on vast GPU clusters to inference on billions of mobile and edge devices.
  • Hardware Heterogeneity: It must be designed for a future with diverse hardware, including different vendors and Meta's own custom silicon.
• 3:50 Announcing a New Suite of Tools: Spisak officially announces a new, multi-layered stack of PyTorch-native tools, detailed in six new blog posts on pytorch.org.
• 5:22 Executor 1.0 for Edge AI: Announcing the 1.0 General Availability (GA) release of Executor, Meta's end-to-end framework for running AI on-device, from wearables to phones. It is supported by a strong ecosystem of partners like Qualcomm and ARM.
• 5:57 Torch Forge for Reinforcement Learning: A new project designed to accelerate research in reinforcement learning (RL) and agentic AI by abstracting away complex infrastructure.
• 6:21 Monarch for Distributed Training: A distributed orchestration framework that uses a single controller. It is presented as a more PyTorch-native and modern approach to large-scale training than the traditional PyTorch Distributed module.
• 6:41 Helium for High-Performance Kernels: A new project that allows developers to write high-performance kernels with significantly fewer lines of code (e.g., 30 lines vs. 120 in Triton) while achieving or exceeding roofline performance through built-in autotuning.
• 7:05 TorchComms for Unified Communications: A communications collectives library created to address hardware fragmentation. It aims to unify communication layers across different hardware (NVIDIA, AMD, custom silicon) into a single, robust, and fault-tolerant library.
• 7:41 Call to Action: Spisak emphasizes that these projects are in their early stages and invites the community to collaborate. He also teases another surprise for his keynote the following day.