The TorchTitan team landed massive improvements in the past year without losing its minimal, extensible ethos. Enhancements include support for SOTA models, dramatic improvements of out-the-box performance at scale, RL support, and new communication features. Read our latest blog to learn more and to get started: pytorch.org/blog/torchtitan-…

Congratulations to the 2026 PyTorch Contributor Award winners, recognized today at #PyTorchCon 2026 for their innovation, collaboration, and impact across all PyTorch Foundation projects including PyTorch, vLLM, DeepSpeed, Ray, Helion, and Safetensors. Outstanding Ambassador: Jiho Kim Newcomer: Jiangyun Zhu (@jiangyun_zhu), Johannes Lennart Zeiser Knowledge Enhancer: Abdulsalam Yazid Bande (@abdulsalambande), Suman Debnath (@_sumand) Bug Buster: Guilherme Vieira Leobas, Vishal Goyal Review Powerhouse: Wentao Ye Catalyst: Jiahao Tan, Harry Mellor (@hmellor_) Innovator: Masahiro Tanaka, Elfie Guo, Xinan Lin Ecosystem Champion: Guangye Yu Superhero: Thomas Ortner Nominees: Chandra Mohan Singh Negi, Arun Bhandari, Masahiro Hiramori, Luka Govedič, Irakli Salia, Jiahao Chen, Kunshang Ji, Rohit Kumar Tiwari, Subin George, Xunzhuo Liu, and Aaron Gokaslan. Thank you to Cara Delia (@RedHat) and Chris Gottbrath (Gottbrath Technologies) for presenting the awards, and to Red Hat (@RedHat) for sponsoring them. And thank you to our 2026 Awards Committee: Chris Gottbrath (Gottbrath Technologies), Chris Hoge (@nvidia), Ashok Emani (@intel), Piotr Bialecki (@nvidia), Yashasvi Misra (@CloudNativeFdn), Sudhir Dharanendraiah (@RedHat), Zesheng Zong (@Huawei), Jiawei Li, and Kaichao You (@inferact).

Wednesday's Demo Theater at #PyTorchCon features 13 presentations on training, inference, hardware enablement, and agentic workflows. All sessions are in the Community Expo on the Concourse Level. 10:40–10:50 AM: Maen Suleiman (@awscloud) demos reinforcement learning on Trainium through TorchNeuron, the native PyTorch backend for AWS Trainium, with TorchTitan training, vLLM-Neuron generating rollouts, Monarch orchestrating, and TorchStore syncing weights. 10:55–11:05 AM: Ali Taha (@baseten) breaks down how Wan 2.2 generates a five-second video in under three seconds, a 53.6x improvement over baseline, using timestep distillation, custom kernel engineering, and four-bit quantization. 12:20–12:30 PM: Rain Jiang and Kevin Li (@verdacloud) demo WideEP for expert-parallel MoE inference on an NVIDIA NVL72 system, using a Rust-based frontend to run and monitor the workload. 12:35–12:45 PM: Tristan Webb (@dMatrix_AI) presents dmx-compressor, a PyTorch 2.0 based quantization framework that gives ML engineers numerical reference output to validate models running on ASIC hardware. 12:45–12:55 PM: Lin She (@clockworkio) demos surviving link and node failures in PyTorch training live, with LinkPass and TorchPass keeping the job running on the same training step. 1:00–1:10 PM: Ashok Emani and Qiacheng Li (@intel) show AI Playground Home Agent, which runs PyTorch AI workflows on an Intel AI PC and lets you use them from your phone. 1:15–1:25 PM: Zesheng Zong (@Huawei) presents CRCR (Cross-Repository CI Relay) for joint testing between PyTorch pull requests and accelerator backends, catching regressions at the PR phase. 1:30–1:40 PM: FNU Harsh and Sarah Ahmed (@LightningAI) demo an agent reproducing ML research on Lightning, showing where it succeeds autonomously and where it needs intervention. 1:45–1:55 PM: George Novack (@inferact) introduces TPU megakernels, with Kimi K3 megakernels running inside vLLM-TorchTPU. 2:00–2:10 PM: Jay Dawani (Lemurian Labs) examines why heterogeneous computing is so hard and what a hardware-agnostic stack has to get right. 2:15–2:25 PM: Minwook Ahn (@RebellionsAI) shares how extending PyTorch's device abstraction to support device-persistent tensors minimizes data movement between host and device. 2:30–2:40 PM: Jesse Treger (HPE) introduces OpenCLIC, an open inference communication framework for connectionless networks that is being integrated with vLLM. 3:55–4:05 PM: Mengjin Yan and Joshua Lee (@anyscalecompute) cover evolving Ray Core for post-training at scale, from scaling Ray's scheduling to 10,000-node clusters to moving PyTorch tensors directly between GPUs with Ray Direct Transport. View the Demo Theater schedule: events.linuxfoundation.org/p…

@marksaroufim Fewer training steps don’t necessarily mean faster training. At #PyTorchCon 2026 (@coreauto) examines the linear algebra kernels behind better optimizers. Mark focuses on QR, symmetric eigendecomposition, and Cholesky, drawing lessons from GPU MODE’s linear algebra competition. He explains how sequential dependencies constrain GPU execution, and how implementations expose parallel work, overlap independent stages, and combine updates to reduce data movement.

FlyDSL, an open source Python kernel DSL, now plugs into PyTorch on @AMD MI350-series GPUs via torch.compile and eager mode. Up to 1.68x geomean scaled GEMM kernel speedups and up to 3.01x end-to-end #vLLM serving speedups. Read more: bit.ly/3W1zwUv

Yiwei Song (@CrusoeAI) shares the vision and progress behind Crusoe’s LLM post-training platform at #PyTorchCon 2026. The goal is to help customers customize open weights models for their business use cases to improve accuracy and reduce inference cost. Yiwei outlines the customization lifecycle: start with a strong base model, inject domain knowledge, align to tasks and preferences, make the model smaller and faster to serve, and deploy it in production. He presents Crusoe Managed RL as a Tinker API-compatible post-training platform, with customers deciding what to learn while Crusoe runs the GPUs. Watch the keynote livestream: events.linuxfoundation.org/p…

Rachad Alao (@cohere) examines how open science can support transparent and secure AI foundation model development at #PyTorchCon 2026. Rachad argues for publishing model architectures, training strategies, parameter choices, and evaluation methodologies, alongside an evidence-based risk framework, independent auditing, continuous monitoring, and dynamic safeguards. He discusses how PyTorch, TorchTitan, and @vllm_project support model development and inference optimization when models are open and accessible, and how access to internal representations enables research in retrieval, reasoning, and multilingual AI.

Zhipeng Wang (@Google) shares technical advances from the @DeepSpeedAI team for large-scale model training at #PyTorchCon 2026. Zhipeng highlights model-systems co-design and optimizations that enable DeepSpeed to deliver high performance across diverse model architectures, workloads, and accelerator platforms. Highlights include AutoSP for automatic sequence parallelism, Muon optimizer support, and AutoTP and AutoEP for Hugging Face models. He discusses the continued development of the DeepSpeed open source community and collaboration with model developers, hardware vendors, and AI infrastructure partners to expand adoption and enable emerging large-scale AI workloads.

Priya Nagpurkar (@IBM) shares how IBM and @RedHat contribute across the PyTorch ecosystem in “From Silicon to Models at Scale” at #PyTorchCon 2026. Priya highlights work bringing IBM Spyre to the PyTorch compiler and vLLM, how AI tools help scale model enablement to thousands of models, and a look ahead at next-generation Spyre hardware. She discusses Red Hat’s contributions to PyTorch and CI/CD infrastructure for out-of-tree accelerators, helping hardware backends keep pace with the community, alongside work with the llm-d community toward efficient distributed inference at scale.

Bill Jia (@GoogleCloud) and Zongwei Zhou (@Meta) introduce TorchTPU, a native PyTorch backend for Google TPUs, and agentic developer workflows at #PyTorchCon 2026. TorchTPU is an open source collaboration between Google and Meta, developed in close partnership with PyTorch core maintainers. It integrates with Hugging Face Transformers for modeling, TorchTitan for training, and vLLM and SGLang for serving. They demonstrate how long-horizon agentic workflows can migrate complex model workloads from GPUs to TPUs. They explore using agents for performance optimization, including quantization, custom kernel generation, and advanced sharding strategies. Watch the keynote livestream: events.linuxfoundation.org/p…

How can LLMs learn from what agents do? At #PyTorchCon 2026, Mazin Gilbert, Executive Director of @AgenticAIFdn, examines the relationship between models and agents. Mazin maps five layers of the agentic stack: hardware infrastructure, software infrastructure, intelligence, the agentic platform, and user applications. He places PyTorch at the intelligence layer and Agentic AI Foundation at the agentic platform layer. He proposes designing the stack in both directions: models support agents, while agent traces, evaluations, environments, and rewards feed back into how models are built. As an open research direction, Mazin describes turning agent evaluations into rewards for training in PyTorch, accounting for safety, task completion, precision, efficiency, and energy use. Watch the keynote livestream: events.linuxfoundation.org/p…

PyTorch Foundation Executive Director Mark Collier @sparkycollier opens Day 2 of #PyTorchCon 2026 with “Building What’s Next in Open Source AI.” Mark looks ahead to broader hardware participation, workloads moving across more environments, and training and customizing frontier models at greater scale. Returning to the vision of open source software to train and serve models on any chip, in any cloud, for any agent, he discusses how AI itself is beginning to change how these systems are built and improved, creating opportunities and challenges for the community. The morning explores agentic AI, workload fungibility, the evolution of inference hardware, scaling large-scale frontier models, contributor recognition, and open science. The focus is on the role open collaboration can play in expanding who can build, contribute, experiment, and shape the future of intelligence, so progress continues to come from many people, projects, and organizations. Watch the keynote livestream: events.linuxfoundation.org/p…

Great in-person sync for the PyTorch TAC Accelerator Integration Working Group on Day 1 of #PyTorchCon! We walked through our recent deliverables: Accelerator Integration Guide torch.accelerator API Additional Compute Platform Pages Ongoing work in flight: Unit Test Refactoring Cross Repository CI Relay (CRCR) Also aligned on short & long-term priorities. The group will keep focusing on backend generality, stability, and release-related work ahead.

Unwind after Day 1 at the Flare Party & Poster Presentations tonight at #PyTorchCon, 6–7:30 PM in the Community Expo (Concourse Level). Explore research, projects, and ideas from the PyTorch community on display, with presenters on hand to discuss their work and answer questions. Good food (sliders, biryani, churros, etc.), cold drinks, even better conversations. Then bring your squad (or join one on the spot) for Pub Trivia from 6:30–7:15 PM in the Demo Theater. No niche tech knowledge required, and the winning team takes home prizes. View the Poster Presentation Schedule: events.linuxfoundation.org/p…

Today's Demo Theater at #PyTorchCon features 14 sponsor presentations on training, inference, compilers, and deployment. Here's what's still ahead this afternoon: 2:00–2:10 PM: IBM presents Spyre’s integration with PyTorch and vLLM, alongside insurance claim processing and document intelligence demos. 2:15–2:25 PM: Datadog examines job-centric observability for Ray training failures. 2:30–2:40 PM: Modal demonstrates low-latency inference and LLM post-training through live Street Fighter III gameplay. 3:55–4:05 PM: Arm connects natural language requests to physical hardware using PyTorch, ExecuTorch, vLLM, and Arm Device Connect. 4:10–4:20 PM: Red Hat tests quantization and speculative decoding under load to examine performance, cost, and quality trade-offs. All sessions are in the Community Expo on the Concourse Level. View the Demo Theater schedule: events.linuxfoundation.org/p…