Welcome to the PyTorch Foundation monthly newsletter!
AI progress is increasingly shaped by the systems that make breakthroughs usable: compilers, kernels, inference engines, distributed training, CI, hardware backends, certification, and the people maintaining all of it in the open. This month’s updates show that work happening across the PyTorch ecosystem.
PyTorch 2.13 brings contributions from 526 contributors, including major progress across attention, memory efficiency, hardware enablement, and developer productivity. We also launched the PyTorch Certified Associate program with Linux Foundation Education, opened nominations for the 2026 PyTorch Foundation Contributor Awards, welcomed Shopify as a Platinum member, and deployed the Cross-Repository CI Relay to improve coordination across PyTorch and out-of-tree backends.
That coordination matters. Open systems learn faster when contributors can see the bottlenecks, improve the shared infrastructure, and make those improvements available to everyone. You can see that same pattern in recent work across SGLang, Helion, vLLM, ExecuTorch, and other PyTorch Foundation-hosted and adjacent projects, from day-zero model support to multi-silicon inference and production kernel optimization.
The pace of innovation is remarkable, but it is not magic. It comes from contributors, maintainers, member companies, researchers, and developers doing the hard systems work together. Thank you for being part of it.
Cheers,
Mark Collier
Executive Director, PyTorch Foundation
Announcements
This release is composed of 3,328 commits from 526 contributors since PyTorch 2.12. We want to sincerely thank our dedicated community for your contributions. Highlights include FlexAttention support on Apple Silicon with up to approximately 12x speedup over SDPA on sparse patterns, the CuTeDSL “Native DSL” backend for key GPU operations, and nn.LinearCrossEntropyLoss to reduce peak GPU memory by up to 4x during large-vocabulary language model training. Read the complete release blog here 👉 PyTorch 2.13 Release Blog
Shopify Joins the PyTorch Foundation as a Platinum Member
Shopify provides essential internet infrastructure for commerce and relies on PyTorch as a key part of applied machine learning across Shopify’s all-in-one platform that makes it easier to start, run, and grow a business. As a Platinum member, Shopify intends to invest in the open source ecosystem it depends on: contributing engineering expertise upstream, sharing what it has learned running machine learning at commerce scale, and helping ensure PyTorch keeps serving the demands of real retail and commerce workloads. Read the full announcement here 👉Shopify Joins the PyTorch Foundation as a Platinum Member
PyTorch Certified Associate Program
Linux Foundation Education and PyTorch Foundation have launched the PyTorch Certified Associate (PTCA), a new certification designed for early-stage practitioners looking to build credibility in AI and machine learning. PTCA validates your ability to work with PyTorch, demonstrating that you understand how models are designed, trained, and used in real-world environments. Enroll today 👉 PyTorch Certified Associate (PTCA)
PyTorch Foundation Contributor Awards
Nominations are now open for the 2026 PyTorch Foundation Contributor Awards! These awards recognize outstanding individuals whose contributions help strengthen PyTorch Foundation-hosted projects, including PyTorch, vLLM, DeepSpeed, Ray, Helion, and Safetensors, as well as the broader community. From technical innovation and documentation to mentorship, advocacy, and community leadership, contributors play a vital role in advancing our mission. 🔥 Submit your nomination by July 17 here 🔥
Upcoming Events
PyTorch 2.13 introduces updates across attention, compilation, distributed training, memory efficiency, Python support, and accelerator platforms. On Wednesday, July 22, 2026, at 11 a.m. PT, PyTorch maintainers and contributors will provide a brief overview of the PyTorch 2.13 release and answer questions from the community live.👉 Register
World Artificial Intelligence Conference (WAIC) 2026, Shanghai, July 17-20
The event features conferences and forums, exhibitions, competitions and awards, application experiences, innovation and incubation programs, and talent and recruitment activities. Pop by the MetaX booth to speak to us at the PyTorch kiosk!
Agentic AI Summit, Berkeley, August 1-2
The Agentic AI 2026 Summit is designed to be the largest and most comprehensive event dedicated to Agentic AI—bringing together leaders across the full stack, from foundation models and agent frameworks to evaluation, infrastructure, and real-world deployment. PyTorch Foundation will have a sponsor booth and CTO Matt White is leading a workshop on “The Open Agentic Stack: Building the Future of AI Systems with Open Source, Open Standards and Composability” on August 1st at 3.50pm.
Community Meetup for PyTorch: Nanshan, Shenzhen, August 15
Meet other PyTorch developers in Nanshan to discuss all things PyTorch. To sign up, send the message “Nanshan” to WeChat subscription account “PyTorch人工智能” to get the registration method.
PyTorch Conference China 2026, Shanghai, September 8-9
The Schedule is now live! View the complete schedule here and register here. Sessions will cover everything from how AI Agents are rewriting the OpenStack/K8s playbook to The State of the PyTorch Ecosystem in 2026: Global Trends and China’s Rising Role. Check out all the session highlights here.
PyTorch Conference North America 2026, San Jose, October 20-21
The schedule will go live next week. Make sure to check it out and register for your early bird ticket before July 31st!
Recent Events
Executorch Hackathon, San Francisco, June 27-28
The aim of the hackathon was to build and optimize real-time AI applications that run directly on Snapdragon-powered mobile devices using ExecuTorch. In total, 372 participants from 65 teams collaborated to develop a variety of impressive projects based on 22 AI applications. Check out the recap blog here 👉 Building the Future of On-Device AI at the ExecuTorch Hackathon
PLDI 2026 AI Summit, Boulder, CO, June 15-19
The Workshop on “Writing Performance-Portable Kernels Simplified with Helion” provided compiler researchers, kernel authors, and ML systems engineers the opportunity to dive deep into the technology. This was an entirely interactive session where attendees got to write, autotune, and run real Helion kernels live.
In the News
- Pytorch: the software layer underpinning Europe’s AI ambitions, Tech EU
- Shopify joins PyTorch Foundation as Platinum member, ITBrief United States
- Will PyTorch certification reset the AI talent benchmark for enterprises, Futurum Group
Latest Blogs
Towards Free Normalization: Fusing Normalization into GEMM and Attention Kernels
In this blog post, we present various novel kernel fusion techniques for common normalization ops like LayerNorm and RMSNorm, which provide significant speedup by reducing the memory-IO overhead of these highly memory-bound kernels.
Bringing PyTorch Monarch to AMD GPUs: Single-Controller Distributed Training on ROCm
Here we explore the architecture of PyTorch Monarch, walk through the engineering effort required to port Monarch’s GPU runtime and distributed communication stack to ROCm, and demonstrate how the system recovers from node failures without halting the entire training job.
Understanding PyTorch’s Test Infrastructure
PyTorch tests are often generated dynamically across devices and dtypes, which is why test names in CI may look different from the class and method names in the source file. This post explains how device-generic tests, OpInfos, instantiate_device_type_tests(), and CI sharding fit together, and how contributors can run PyTorch tests more effectively.
Miles: A PyTorch-Native Stack for Large-Scale LLM RL Post-Training
Miles is RadixArk’s open source framework for large-scale LLM RL post-training. It composes SGLang for rollout, NVIDIA Megatron-LM for training, Ray orchestration, and PyTorch-native extensibility behind a small, pluggable trainer, with unified low-precision recipes, MoE-aware rollout/training alignment, fast NVIDIA NCCL/RDMA weight synchronization, observability, and fault tolerance built in — making frontier-scale LLM RL easier to build, reproduce, and operate.
Introducing Cross-Repository CI Relay: Scalable CI for PyTorch’s Out-of-Tree Backends
PyTorch now has a Cross-Repository CI Relay (CRCR) that automatically triggers and tracks CI in downstream repositories whenever a PR is opened or a commit is pushed against pytorch/pytorch. Results flow back to the PyTorch CI HUD, giving maintainers a single dashboard for both in-tree and cross-repository CI health.
TokenSpeed-Kernel: Portable APIs and High-Performance Kernels for Multi-Silicon LLM Inference
The TokenSpeed-kernel is an open-source subsystem designed to solve backend complexity in LLM inference. This blog provides a technical breakdown of the TokenSpeed-kernel and how it helps developers work with high-performance kernels for multi-silicon LLM inference.
Serving DeepSeek-V4 on GB300 with SGLang: 5x Higher Throughput at the Same Interactivity Since Day-0
DeepSeek-V4 support was live in SGLang on Day-0, but the Day-0 stack was only the starting point. Since launch, a set of kernel, runtime, and hardening improvements have been coordinated, including MHC fusion and token-bucket prewarm, KV Compression V2, W4A4 MegaMoE, stronger SWA budgeting and eviction behavior. In this blog we look at the performance results.
From Minutes to Seconds: LLM-Guided Autotuning for Helion Kernels
Helion relies on autotuning for performance. Currently Helion searches utilize the Likelihood-Free Bayesian Optimization (LFBO) to find the most performant configs. This blog introduces LLM-guided autotuner as a practical alternative to faster kernel tuning at production quality.
Portable vLLM Model Inference Kernels in Helion
Helion kernels were integrated into vLLM for FP8 inference using Qwen3 models and evaluated across NVIDIA H100 and B200 GPUs. Learn how Helion provides a productive PyTorch-native workflow for developing fused GPU kernels while delivering performance improvements for quantization, normalization, and fusion-heavy inference kernels.
PyTorch Meetup Singapore: A milestone in APAC
Eighty engineers, researchers, and community builders gathered for the inaugural PyTorch Meetup Singapore. In this blog, we take a closer look at the topics discussed: inference, distributed training, continuous integration, community governance, and the broader question of technological sovereignty.
PyTorch Foundation Hosted Project Updates
- vLLM Conference 2026, San Francisco, August 24–26
The first-ever vLLM Conference at Ray Summit, is coming to San Francisco. This event features sessions on where the vLLM roadmap is headed, getting the most out of accelerators, and serving inference in production. There will be speakers from PyTorch, Inferact, NVIDIA, AMD, TPU, Anyscale, and more. Register and meet the vLLM community in person.