PyTorch Conference North America 2026 聚焦開放智能與 PyTorch 生態更新

PyTorch Conference North America 2026 brought 3,500+ attendees to San Jose on October 20–21 for two days focused on accelerating open intelligence and exploring the impact of PyTorch Foundation projects like PyTorch, vLLM, DeepSpeed, Ray, Helion, and Safetensors.

PyTorch Foundation Executive Director, Mark Collier, set the stage for the entire event to demonstrate how the PyTorch Foundation is building the open source software to train and serve models on any chip, in any cloud, for any agent.

Accelerating Open Intelligence Together

Mark Collier’s opening keynotes connected work across research, models, frameworks, infrastructure, hardware, and production. Day 1 examined how shared advances across those layers help the community move faster together. Day 2 looked ahead to broader hardware participation, workloads moving across more environments, and training and customizing frontier models at greater scale.

Recent milestones across the PyTorch ecosystem highlight how AI itself is joining the learning loop to improve code, kernels, and research workflows. Announcements and updates published during the conference and in the preceding days covered the following work.

Native PyTorch on Google TPUs

TorchTPU is an open source collaboration between Google and Meta that introduces a native and principled PyTorch backend for Google TPUs. It supports eager execution and torch.compile, with validated integrations covering TorchTitan for training, vLLM and SGLang for inference, and Ray for orchestration. Workload migration can require only a device change, while peak performance can require additional changes to shapes, sharding, and operations.

Read Introducing TorchTPU.

TorchNeuron for AWS Trainium

TorchNeuron is an open source native PyTorch backend for AWS Trainium (and corresponding Amazon EC2 instances: Trn2 and Trn3), enabling AI researchers and practitioners to experiment and run PyTorch workloads with a single configuration change. With eager execution, torch.compile, and native support for PyTorch APIs (DTensor, torch.distributed) and libraries like TorchTitan, TorchNeuron enables workflows from research experimentation to production deployment. The public beta launch under Apache 2.0 was on October 20th at PyTorch Conference.

Read Introducing TorchNeuron.

Compiler Development for IBM Spyre and AMD GPUs

IBM’s “From KTIR to SpyreCode” explains how torch-spyre translates Inductor’s decisions about work division and tensor placement into executable programs. An open source Dataflow Scheduler builds pipelines for data movement and computation before device-specific stages produce SpyreCode. The October 21 FlyDSL update also described an optional backend for AMD MI350-series GPUs, connecting eligible operations to optimized kernels through existing PyTorch APIs while retaining existing implementations for unsupported workloads.

Read From KTIR to SpyreCode.

TorchTitan and TitanRL

TorchTitan expanded model support across its 5D parallelism stack and introduced reinforcement learning through TitanRL. TitanRL keeps the same model implementation across pretraining, post-training, and execution inside vLLM, with asynchronous rollouts and multiple forms of parallelism for training and generation. The update also covers performance improvements, FlexShard, graph-based training, distributed communication, and checkpointing.

Read the TorchTitan update.

Helion and NVIDIA CompileIQ

Helion’s integration with NVIDIA CompileIQ adds compiler controls to the kernel autotuning search. Developers can evaluate Advanced Control Files alongside kernel configurations such as tile sizes and warp counts, tune offline, and pin the selected configuration. Compiler defaults remain a candidate in the search.

Read about Helion and NVIDIA CompileIQ.

ExecuTorch on Arm Cortex-M CPUs

ExecuTorch now provides a direct path from PyTorch training to on-device inference. The Cortex-M backend enables PyTorch models to run directly on Arm Cortex-M CPUs with Arm CMSIS-NN acceleration. It supports all capability tiers (Scalar, DSP, and MVE/Helium) from a single workflow.

Read Bringing ExecuTorch to Arm Cortex-M CPUs.

Nine New Members Join PyTorch Foundation

New Platinum members include Nebius, Nscale and Shopify, deepening the PyTorch Foundation’s role as the intelligence layer for open source AI development.

The Foundation welcomes the following new members:

Platinum: Nebius, Nscale, and Shopify Inc.
Silver: LightSeek Foundation and Lablup Inc.
Associate: SKOPOS INDIA FOUNDATION, Siksha ‘O’ Anusandhan, Messiah University, and Valencia College.

AI innovation now happens across all stack layers simultaneously, with research, models, infrastructure, hardware, and production continuously driving each other forward. The PyTorch Foundation serves as the central hub of this interconnected ecosystem, uniting researchers, developers, and companies building open AI.

“Open systems learn faster when more people, projects, and organizations are working together, creating compounding gains,” said Mark Collier, Executive Director of PyTorch Foundation. “Open source ecosystems let breakthroughs travel between organizations instead of staying locked inside one company.”

Nebius has also long supported vLLM with compute for development, testing, and benchmarking, including multi-node work on large models such as DeepSeek R1.

Nscale first came to the PyTorch Foundation’s project ecosystem through Ray, the open-source framework for distributed computing and scaling AI workloads. Following Nscale’s planned acquisition of Anyscale, the company behind Ray, Nscale has committed to continuing investment in Ray’s development and its contributor community, building on Ray’s governance under the Foundation since the project joined in October 2025.

“Most infrastructure providers treat the framework and the data center as two separate problems. We treat them as one. PyTorch has remained the base layer of AI because it kept the Pythonic style researchers know, while getting serious about performance. Ray, also part of the PyTorch Foundation, does something different: getting that work to run across thousands of GPUs. With Anyscale, the creators of Ray, set to join Nscale, we’re not just using that ecosystem, we’re building the infrastructure underneath it. Joining the PyTorch Foundation as a Platinum member is us backing that directly, not just benefiting from it,” said Nick Jones, VP of Cloud Native Platform Engineering, Nscale

Shopify’s work featured in Cody Mazza-Anthony’s keynote and an accompanying case study. Using a continual learning loop with PyTorch and vLLM, Shopify compresses production failures into model weights every day, beating frontier-model quality, and cutting serving costs by 96%.

Day 0: Training, Workshops, and Co-Located Events

Day 0 programming on October 19 included:

PyTorch, Powering the Enterprise with Red Hat, NVIDIA, and IBM
LF AI & Data Summit 2026
PyTorch Associate Training
Jupyter Day
From Academia to Adoption Workshop
Arm Create with Google, Google DeepMind, Nota, and Renesas
Fine-Tune LLMs on Amazon SageMaker and Run Them On-Device with GenieX

The PyTorch Associate Training offered an in-person, instructor-led program focused on the foundational and practical skills covered by the PyTorch Certified Associate (PTCA) exam. Through lectures, demonstrations, guided labs, and assessments, the training was designed to help participants build and optimize real-world models from scratch. Course completion included a voucher for the certification exam.

Thank you to lead trainer Faradawn Yang, content creator Sahdev Zala, and assistant trainers Allison Ding, Krishna Kalyan, Santosh Bahir, Lakshay Garg, and Ian Wang.

Hosted by the PyTorch Foundation OSPO & Academic Outreach Working Group, the From Academia to Adoption Workshop focused on taking promising PyTorch projects beyond a single paper, experiment, or lab. The community event explored how research software can become reliable, reusable tools that others can adopt, contribute to, and build upon.

Two Days of Keynotes

Tuesday, October 20

Accelerating Open Intelligence Together: Mark Collier (PyTorch Foundation)
Shopify’s Continual Learning Loop with PyTorch and vLLM: Cody Mazza-Anthony (Shopify)
PyTorch Updates: Alban Desmaison and Edward Yang (Meta)
MAX and Mojo: The Best Way to Extend PyTorch: Chris Lattner (Qualcomm)
Ray’s Evolution to Support AI Workloads: Ion Stoica (UC Berkeley / Anyscale)
Trainium’s Journey to Native PyTorch: Robert Hundt (Amazon)
Beyond Brute Force: The Era of Adaptive Intelligence: Sara Hooker (Adaption)
PyTorch at Meta – From Internal Workloads to Multi-Silicon Ecosystem: Jana van Greunen (Meta)
vLLM Update: Scaling Open Frontier Inference Infrastructure: Simon Mo (Inferact)
Open Source Inference at Production Scale: Carlos Costa (IBM), Abdullah Gharaibeh (Google Cloud), and Stephen Watt (Red Hat)
PyTorch Performance for Agents on Vera Rubin: Ujval Kapasi (NVIDIA)

Wednesday, October 21

Building What’s Next in Open Source AI: Mark Collier (PyTorch Foundation)
From AI to Agentic AI and Back: Mazin Gilbert (Agentic AI Foundation)
Workload Fungibility
in the Agentic Era: Bringing Native PyTorch to Cloud TPUs: Bill Jia (Google Cloud) and Zongwei Zhou (Meta)
From Silicon to Models at Scale: Priya Nagpurkar (IBM)
Contributor Awards – Sponsored by Red Hat: presented by Cara Delia (Red Hat) and Chris Gottbrath
Scaling Large Frontier Models Training with DeepSpeed: Zhipeng Wang (Google)
Open Science and AI Foundation Models: Allies in the Pursuit of Knowledge: Rachad Alao (Cohere)
AI Model Customization Platform: Yiwei Song (Crusoe)
Linear Algebra for the Age of Research: Mark Saroufim (Core Automation)

Technical Sessions and Community Expo

This year’s impressive program featured more than 100 sessions spanning foundational concepts and core framework work to training, inference, applications, kernel engineering, and responsible AI. Each track highlighted a distinct part of how PyTorch Foundation projects are built, extended, deployed, and used in the real world. View the full schedule and stay tuned to the PyTorch YouTube channel where all sessions will be published after the conference.

PyTorchCon community expo Beyond technical sessions, the conference also fostered community collaboration with engaging onsite experiences in Community Expo including Birds of a Feather (BoFs) discussions, live showcases at the Demo Theater, direct Q&A with core maintainers during Meet the Developers, and dedicated networking events like the Women & Non-Binary in PyTorch Lunch. Attendees lined up to pick up an autographed copy of Inference Engineering at the book signing with Philip Kiely and there were endless opportunities to connect and innovate. 144 poster sessions showcased a broad range of community-led innovations spanning core PyTorch developments, kernel engineering, model training, on-device inference, and responsible AI. Key highlights include performance optimization techniques for LLMs on diverse hardware architectures, efficient on-device deployment frameworks like ExecuTorch, and novel agentic workflows.

Celebrating the 2026 PyTorch Contributor Award Winners

The 2026 PyTorch Contributor Awards recognized innovation, collaboration, and impact across PyTorch Foundation projects: PyTorch, vLLM, DeepSpeed, Ray, Helion, and Safetensors.

Congratulations to the winners:

Outstanding Ambassador: Jiho Kim
Newcomer: Jiangyun Zhu and Johannes Zeiser
Knowledge Enhancer: Abdulsalam Yazid Bande and Suman Debnath
Bug Buster: Guilherme Vieira Leobas and Vishal Goyal
Review Powerhouse: Wentao Ye
Catalyst: Jiahao Tan and Harry Mellor
Innovator: Masahiro Tanaka, Elfie Guo, and Xinan Lin
Ecosystem Champion: Guangye Yu
Superhero: Thomas Ortner

Congratulations to 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 of Red Hat and Chris Gottbrath for presenting the awards, and to Red Hat for sponsoring them.

We also thank the 2026 Awards Committee: Chris Hoge (NVIDIA), Chris Gottbrath, Ashok Emani (Intel), Piotr Bialecki (NVIDIA), Yashasvi Misra, Sudhir Dharanendraiah (Red Hat), Zesheng Zong (Huawei), Jiawei Li, and Kaichao You (Inferact).

Flare Party and AI Community Bash with De La Soul

The Community Expo brought attendees together with sponsors and exhibitors at their booths to discuss work across the open source AI stack. Conversations continued after Day 1 at Tuesday’s Flare Party, with food, drinks, and Pub Trivia in the Demo Theater.

On Wednesday, the AI Community Bash brought the AGNTCon + MCPCon and PyTorch communities together for an evening of food, drinks, great company, and a live concert from legendary hip-hop group DE LA SOUL.

Thank You to Our Program Chairs, Committee, and Sponsors

Thank you to Program Chairs Jeff Daily (AMD), Alban Desmaison (Meta), and Mike Ruberry (NVIDIA), and to the full Program Committee for contributing their time and expertise.

Meet the program chairs and committee.

Thank you to our Diamond sponsors: AMD, AWS, Crusoe, Google, IBM, Meta, NVIDIA, and Qualcomm.

Thank you to our Platinum sponsors: Anyscale, Arm, Baseten, Clockwork.io, Huawei, Inferact, Intel, Lightning AI, Nscale, Red Hat, and Verda.

We appreciate every sponsor’s support. Visit the conference website for the full sponsor listing.

Looking Ahead to 2027

Thank you to the contributors, maintainers, speakers, attendees, and member organizations who came together in San Jose to accelerate open intelligence.

Stay tuned for session slides and recordings, and save the date: PyTorch Conference returns to San Jose, October 6–7, 2027!

The post PyTorch Conference North America 2026: Accelerating Open Intelligence Together appeared first on PyTorch.