PyTorch

What belongs at Open Source AI Week? There is plenty of room in the model. 💡 Events can focus on open technologies throughout the AI ecosystem, including: 💻 Open source software and hardware for AI development 📐 Open standards 📊 Open data related to AI 🧪 Open benchmarks 🤝 Community collaboration and inclusive innovation Eligible events must take place in the Bay Area between October 16-25. Your event will join a regional lineup anchored by PyTorch Conference North America, October 20-21, and AGNTCon + MCPCon North America, October 22-23, both in San Jose. Have an event that fits? Submit it by October 15 at 11:59 PM PST: https://bit.ly/4ipYZ39 View the week: https://bit.ly/3KjXQv2 #OpenSourceAIWeek #PyTorch #PyTorchFoundation #PyTorchCon #AGNTCon #MCPCon #AAIF #FutureOfAI #AI #GenAI #MachineLearning #ML #DeepLearning #OpenSource #OpenSourceSoftware #OpenSourceDevelopment #OpenSourceCommunity #OSS #LinuxFoundation #events #linux

How do you keep vLLM moving at the speed of light without excluding users who run diverse models on diverse hardware? In a new PyTorch Foundation blog, contributors from IBM, Meta, and Hugging Face introduce hardware-agnostic layers designed to balance frontier performance with portability, helping ensure vLLM continues to meet the needs of the broader open-source ecosystem. Read the blog to learn more: https://lnkd.in/e-rUheK2 Thomas Parnell, Thomas Ortner, Harry Mellor, Richard Zou

Where research meets real-world application. 🌎 At PyTorch Conference North America, hear from the researchers and practitioners advancing PyTorch and putting those advances to work. See how ideas move from research into practical implementation and how what gets learned in practice helps shape what comes next. Join us in San Jose, October 20-21: https://bit.ly/4sh3DSw #PyTorchCon #PyTorch #PyTorchFoundation #FutureOfAI #AI #GenAI #MachineLearning #ML #DeepLearning #OpenSource #OpenSourceSoftware #OpenSourceDevelopment #OpenSourceCommunity #OSS #LinuxFoundation #events #linux

PyTorch releases are no longer just about shipping PyTorch binaries. Release engineering now coordinates PyTorch, Triton, and vLLM as part of a broader ecosystem. At PyTorch Conference North America 2026, Andrey Talman (Meta) will discuss how the PyTorch release process is being modernized, including the use of AI agentic workflows to analyze failures and the work required to release projects together with reliability and performance in mind. Hear more from Andrey about what he’ll cover at the conference. 🔗 Register: https://lnkd.in/geC5_z9a #PyTorchCon

Some of the best conversations happen when you bring curious people together in the same room. Last week, we welcomed the PyTorch London community to our London office for an evening of sharing ideas, perspectives and plenty of thoughtful discussion. After food and drinks on the rooftop, Helen Byrne welcomed everyone before talks from Arsalan Uddin, Andrew Fitzgibbon, and Akshat Tripathi brought different insights to the room. Big thank you to Sylvain Viguier for bringing the community together and putting so much care into making the meet up happen, alongside Idalecio Rosa, Seán Comerford, Luke Hudlass-Galley, Mark Rankilor, Michele Taroni, Gareth Hubbard and their support for the evening. We enjoy opening our doors for moments like these: a chance to share some of the thinking from our research teams, learn from others, and give the wider community space to exchange ideas. Thank you to everyone who joined us.

Frontier models are the fastest way to launch an AI product, but what happens when usage scales? They don't learn from production failures on their own, and serving them can become incredibly slow and expensive. In our latest case study, Shopify shares how they built a continual learning loop using PyTorch and vLLM for their GraphQL agent, turning everyday production failures directly into model weight improvements. Key wins from their self-healing pipeline include: - 96% reduction in serving costs compared to relying on frontier models at scale. - Higher output quality achieved through supervised fine-tuning and GRPO directly against their calibrated quality rubrics. - 38% drop in end-to-end latency by compressing long, static system prompts into a short sequence of learned gist tokens. - Automated continuous learning where a panel of reasoning models critiques failures and generates new, successful training trajectories. Instead of just accumulating discrete artifact updates like prompt edits or complex routing rules, Shopify is compressing their production experience directly into the continuous space of the model's weights. Read the full breakdown by Cody Mazza-Anthony and Andrew McNamara to learn more about they define quality, calibrate judges, and distribute training across GPUs: https://lnkd.in/e9Dtx7au Plus, to dig in even deeper, don't miss Cody Mazza-Anthony's keynote at PyTorch Conference North America on how small models can outperform frontier models on well-scoped tasks at a fraction of the cost, with lower latency and higher throughput. Get your ticket for PyTorch Conference NA (October 20-21 in San Jose, CA): https://hubs.la/Q04v8Lq_0 #PyTorchCon

Every mature systems project eventually needs a teaching version. TinyTorch is a free, open source curriculum where you build a working machine learning framework from scratch, tensors through transformers, using PyTorch’s own API in pure Python. It requires no GPU, runs on a 4 GB laptop, and covers 20 hands-on modules designed to give developers, students, and engineers a complete mental model of PyTorch internals. Read the full technical breakdown here: https://lnkd.in/e9FAgK68 Vijay Janapa Reddi, Andrea Mattia Garavagno

RL post-training has become a critical stage in modern LLM development, but deploying an end-to-end pipeline requires much more than running individual kernels efficiently. Systems must coordinate distributed training, rollout generation, and continuous weight synchronization across multiple software stacks while maintaining correctness and performance. Joy Song with AMD will present a poster at PyTorch Conference North America on enabling the open source vime RL post-training framework on AMD Instinct GPUs using ROCm. vime pairs Megatron with vLLM for high throughput rollout generation, forming a tightly coupled loop where model weights are continuously synchronized between the two systems. Joy will focus on the engineering challenges this brings up, including correctness issues that emerge only during integrated training and rollout execution and how systematic debugging can strengthen open source PyTorch infrastructure on AMD GPUs. Register for PyTorchCon NA at: https://hubs.la/Q04v4SL60 #PyTorchCon

Sometimes seeing is believing. Sometimes seeing is benchmarking. ⚡ Stop by the Demo Theater in the Community Expo at #PyTorchCon North America for compact sessions showing how PyTorch technology performs beyond the presentation deck. The lineup includes live looks at: 🔥 PyTorch on TPUs and Trainium 📱 AI spanning cloud, edge, and embedded devices 🛡️ Resilient distributed training ⚙️ Low-precision training and inference 🚀 Model deployment and hardware portability Add a few demos to your agenda for October 20-21 in San Jose: https://bit.ly/4hb0ekq Register: https://bit.ly/4sh3DSw #PyTorchCon #PyTorch #PyTorchFoundation #FutureOfAI #AI #GenAI #MachineLearning #ML #DeepLearning #OpenSource #OpenSourceSoftware #OpenSourceDevelopment #OpenSourceCommunity #OSS #LinuxFoundation #events #linux

Join Valérian Rey from SimplexLab at PyTorch Conference North America 2026 as he breaks down multi-objective optimization. He will share practical techniques for training models with multiple loss functions simultaneously and demonstrate how to easily implement these workflows using the TorchJD library. Whether you are optimizing multi-task architectures or managing complex loss tradeoffs, this session will provide concrete steps to streamline your pipeline. Join the open source AI community in San Jose, October 20-21 : https://hubs.la/Q04v4SL60 #PyTorchCon