PyTorch Newsletter

March, 2026

March 25, 2026No Comments

March 2026 Newsletter Header

Welcome to the March edition of the PyTorch Foundation Newsletter. Inside: the release of PyTorch 2.11 and our upcoming Live Q&A, the schedule for PyTorch Conference Europe, registration for the KubeCon + CloudNativeCon + OpenInfra Summit + PyTorch Conference China 2026, new member announcements, and more.

Announcements

PyTorch 2.11 Release
PyTorch 2.11 Release Now Live: We are excited to announce PyTorch 2.11, featuring 2,723 commits from 432 contributors. Key updates include differentiable collectives for distributed training, a FlashAttention-4 backend for FlexAttention on Hopper/Blackwell, and expanded MPS operator support for Apple Silicon. Read the release notes >>
PyTorch Conference China 2026
KubeCon + CloudNativeCon + OpenInfra Summit + PyTorch Conference China 2026 CFP & Registration Now Open: Both the Call for Proposals and registration are now open for the co-located event taking place 8-9 September in Shanghai. Submit to speak >>
New PyTorch Members
PyTorch Foundation Announces New Members as Agentic AI Demand Grows: The Foundation recently welcomed nine new members, including Clockwork.io, Emmi AI, Nota AI, and yasp.ai. This expansion supports the rapid growth of open source AI infrastructure.

Upcoming Events

OpenEnv AI Hackathon (India): Meta, Hugging Face, and PyTorch present India’s first OpenEnv AI Hackathon. Build reinforcement learning environments using Meta’s open-source RL framework for a chance to win from a $30,000 prize pool. Round 1 begins online March 25. Register by April 3 >>
Optimized AI Conference: Join PyTorch Foundation in Atlanta (March 30–31) for the Optimized AI Conference. We’re proud to be a Community Partner; visit our booth to connect with the team. Learn more >>
PyTorch 2.11 Webinar
PyTorch 2.11 Release Live Q&A: Join Andrey Talman and Nikita Shulga on March 31 at 10 am PT for a technical deep dive into the 2.11 release. Topics include differentiable collectives, RNN/LSTM GPU export, and XPU Graph. Register here >>
PyTorch Conference Europe 2026
PyTorch Conference Europe 2026: The schedule is live for Paris, April 7–8. Featuring 95+ technical sessions across eight tracks, covering audio modality for LLMs, C++ wrapper modes, and responsible AI engineering. Register now >>
GPU MODE IRL Hackathon (Paris): Following PyTorchCon EU on April 9, this technical sprint focuses on ML systems optimization and LLM speedruns. Participants get access to NVIDIA Blackwell Ultra clusters to attempt pre-training an LLM from scratch. Learn more >>

Recent Events

PyTorch Day India Recap
PyTorch Day India 2026 Recap: The builder-focused event in Bengaluru brought 460 in-person attendees together for technical talks and interactive discussions to foster collaboration within the Indian open source AI community.
PyTorch at NVIDIA GTC
PyTorch at NVIDIA GTC 2026: Maintainers showcased high-performance AI integration in San Jose. Featured demos included the new Helion framework for kernel authoring and ExecuTorch deployments.
PyTorch Helion Hackathon: On March 14, developers gathered in San Francisco for a hands-on event focused on GPU kernel authoring and performance optimization using the Helion framework.

Member Round Up

Intel Core Ultra AI PC
Intel: Unlocking AI Capabilities on Intel® Core™ Ultra Series 3 Processors with PyTorch 2.10 — Highlights advancements in the XPU backend, including TorchAO integration for seamless int4 quantization.
DeepSpeed Multimodal Efficiency
Snowflake: Enhancing Multimodal Training and Memory Efficiency — Updates enable a 30% speedup and 40% reduction in peak memory via DeepSpeed.
KernelAgent
Meta: KernelAgent: Hardware-Guided GPU Kernel Optimization — Introduces an agentic system that achieves a 1.56x speedup over default torch.compile.
MXFP8 Training
Meta: MXFP8 Training for MoEs — Demonstrates a +30.2% speedup for Llama4 Scout training on GB200 clusters using MXFP8.
Subscribe to the PyTorch Newsletter
Get updates directly to your inbox: https://pytorch.org/newsletter/