PyTorch

The PyTorch Foundation Ambassador Program is open for 2026 nominations. The program supports community leaders who organize events, create technical content, mentor contributors, and support regional PyTorch communities worldwide. We especially encourage applications from contributors across Africa, Latin America, the Middle East, Oceania, Southeast Asia, and Eastern Europe as we continue expanding representation across local PyTorch communities. Ambassadors receive recognition through PyTorch Foundation channels, access to collaboration and training opportunities, support for approved community activities, and complimentary admission to PyTorch Conference. Nominations are open through June 18, 2026. Individuals may nominate themselves or someone else. 🔗 Learn more and apply: pytorch.org/blog/join-the-…

PyTorch 2.12 introduced major updates across compilation, distributed systems, export, graph capture, and accelerator support. Wednesday’s PyTorch 2.12 Release Live Q&A provided a technical overview of the release and answered questions from the community live. Andrey Talman (@Meta), @albanDesmaison (@Meta), and @joespeez (@reflection_ai), moderated by @Chris_AI_HPC, discussed updates including the new device-agnostic torch.accelerator.Graph API, torch.export.save support for microscaling quantization formats, torch.cond capture within CUDA Graphs, and expanded CUDA, ROCm, XPU, MPS, and Arm platform support. This clip highlights part of the discussion around the range of execution workflows supported in PyTorch, including eager mode, torch.compile, and export and graph-based methods, and the different needs of users working across training, inference, and deployment environments. Thank you to everyone who joined live. 🔗 Watch the full video: youtube.com/live/Zn_flVwhH…

Federated Learning Without the Refactoring Overhead The most valuable data is often the least movable. Regulatory boundaries, data sovereignty rules, and organizational risk tolerance routinely prevent centralized aggregation. Meanwhile, sheer data gravity makes even permitted transfers slow, expensive, and fragile at scale. The latest version of NVIDIA FLARE addresses this reality with a Federated Learning (FL) computing runtime that moves the training logic to the data, while raw data stays put. See examples of how to leverage PyTorch in a federated learning system. Read the full post: developer.nvidia.com/blog/federated…

NVIDIA shared a full playlist of its sessions from PyTorch Conference Europe. NVIDIA engineers and developers covered a wide range of topics, including: - Symmetric Memory + NCCL Device APIs - Custom Kernels in torch.compile - Agentic AI & LLM Evaluation frameworks, and more. The clip featured here is from Besmira Nushi's keynote, "The Unbearable Lightness of (Agentic) Evaluations," which explores how agentic evaluation frameworks expose gaps between what models say and what they actually do. Find the full playlist here: youtube.com/playlist?list=…

Read the blog: pytorch.org/blog/pytorch-d…

The PyTorch Docathon 2026 has officially wrapped up, and the results are incredible! Over the past two weeks (May 5–19), our open source community showed up with massive energy and dedication to make PyTorch documentation better for developers everywhere. Here is a snapshot of what we accomplished together: Ecosystem Reach: 260+ registrants and 30+ active participants collaborated globally. Impactful Contributions: 150+ merged pull requests successfully landed across all difficulty levels. Technical Depth: Community members resolved critical issues, added API documentation, and expanded core ExecuTorch documentation. Why does high-quality documentation matter more than ever in the AI era? Clear documentation lowers the barrier to entry for human developers and directly fuels the modern AI ecosystem. LLMs and AI agents increasingly rely on public technical documentation to learn APIs, generate code, and troubleshoot complex workflows. High-quality PyTorch docs help ensure AI-generated guidance is more accurate, up-to-date, and aligned with best practices. 🏆 Celebrating Our Community Leaders: A massive shout-out to the top-performing contributors who led the way and whose work improves the experience for millions of PyTorch users worldwide: First Place: ymrohit Second Place: XAheli, PyDevC, and darknight054 Third Place: JonathanColetti and Kadermiyanyedi Honorable Mentions: AswaniSahoo, Vasanthadithya-mundrathi, Nazim-fad, ozgecinko, kiszk, saurabhkthakur, and spzala. Thank you to everyone who brought code, reviews, and the community momentum!

Expert-led sessions, hands-on learning, and networking. 🤝 #PyTorchCon North America is the place to be this Oct 20-21 in San Jose. Save $400 when you register by July 31! Details: bit.ly/4sh3DSw

Developing a Real-Time Hailstorm Forecasting Model With the goal to provide critical weather observation and nowcasting capabilities for severe weather prediction, Colorado State University used PyTorch-based NVIDIA PhysicsNeMo and trained a specialized version of the NVIDIA Earth-2 StormScope Nowcasting model, specifically for operational severe hailstorm prediction. 🔗 Read the full case study: nvidia.com/en-us/case-stu…

@vikhyatk Thank you for flagging. We removed that sentence and noted the update since pytorch/pytorch#177276 did not land in the 2.12 release

@PyTorch the release notes talk about 'torch._C._mps_loadMetalllib', but it seems like that was reverted? I see it in nightly but not 2.12

PyTorch 2.12 introduces major updates across compilation, export, distributed training, and accelerator support. Highlights include up to 100x faster batched linalg.eigh on CUDA, the new torch.accelerator.Graph API, Microscaling quantization support in torch .export.save, and fused Adagrad. The release includes 2,926 commits from 457 contributors since PyTorch 2.11. Have questions? Join @AndreyTalman (@Meta), @albanDesmaison (@Meta), and @joespeez (@reflection_ai), moderated by @Chris_AI_HPC (@Meta), on May 20 at 10:00 AM PT for a live Q&A covering the release and answering questions from the community. 🔗 Read the release blog and register for the webinar: pytorch.org/blog/pytorch-2… #PyTorch #OpenSourceAI #MachineLearning #AIInfrastructure

Excited to see Shepherd Model Gateway featured in The Linux Foundation’s May 2026 newsletter. SMG explores a Rust-based gateway architecture that moves non-GPU orchestration work out of the inference path, helping improve throughput and efficiency for modern long-context LLM serving systems. Thanks to the LightSeek Foundation and the broader PyTorch ecosystem for the collaboration and support. pytorch.org/blog/lightseek…

I know, these days stars don't mean much, but it's still lovely to see such a nice round number. Thank you everyone for trusting your ML workloads to us, here's to the next 100K stars! (H/t Tianyu for pulling this screenshot)

vLLM and PyTorch worked together to fix a long-standing aarch64 install headache — as of PyTorch 2.11.0, pip install torch on GB200 / GB300 / GH200 just works. 🎉 What changed: PyTorch 2.11.0 now publishes CUDA-enabled aarch64 wheels to the default PyPI index. No more custom --index-url flags. No more transitive dependencies silently swapping your GPU build for the CPU wheel. New users on Grace Hopper and Grace Blackwell systems can follow the standard install instructions and have vLLM work the first time. In our latest blog, @KaichaoYou (co-founder @inferact, Lead Maintainer @vllm_project) shares the full story: 🐛 A 2024 hackathon bug bringing up vLLM on GH200 🔧 vLLM's in-tree workarounds (use_existing_torch.py and [tool.uv] build-isolation passthrough) 🤝 From GitHub issue to PyTorch Foundation TAC discussion 🚀 The fix landing in PyTorch 2.11.0, driven by NVIDIA and PyTorch core. A great example of cross-project collaboration under the PyTorch Foundation umbrella — and a reminder that boring infrastructure wins compound. Read the full story: pytorch.org/blog/vllm-and-… ✍️ : Piotr Bialecki (@nvidia) — @ptrblck_de, Alban Desmaison (@Meta), Andrey Talman (@Meta), Nikita Shulga (@Meta)

Don't miss @DeepSpeedAI virtual office hours on May 26 at 12:00 PM America/New_York to ask questions of @toh_tana member of DeepSpeed TSC & get the latest recent key updates, including AutoSP (sequence parallel), AutoEP (expert parallel), and AutoTP (tensor parallel).

ExecuTorch now has an MLX delegate that runs PyTorch models on Apple Silicon GPUs. It supports LLMs, speech-to-text, and MoE models with quantization via TorchAO. Export with torch.export, run on Metal. Read our latest blog: pytorch.org/blog/running-p…

PyTorch releases include thousands of changes, and our Release Live Q&A gives you direct access to the maintainers and contributors behind them. In this clip from our PyTorch 2.11 webinar, Nikita Shulga explains why quantization support is moving from PyTorch core to TorchAO and what that means for developers. Have questions? Join @AndreyTalman, @albanDesmaison, and @joespeez, moderated by @Chris_AI_HPC, on May 20 at 10:00 AM PT for a live Q&A covering PyTorch 2.12 and answering questions from the community. 🔗 Register now: pytorch.org/event/pytorch-…

Big congrats to the ExecuTorch team. Their paper just won the Best Industry Paper Award at @MLSysConf 2026. 🏆 ExecuTorch is PyTorch's framework for running models on-device, in production today across phones, wearables, desktops, and embedded systems. Backends include Apple, Arm, Qualcomm, MediaTek, Samsung, NXP, Cadence, Intel, NVIDIA, and others. Catch the team Tuesday morning in the Best Paper Session ( AM PT) and at the poster session Thursday evening. Paper: arxiv.org/abs/2605.08195 Discord: discord.gg/Dh43CKSAdc X: x.com/Executorch Project: github.com/pytorch/execut… Schedule: mlsys.org/virtual/2026/c…