PyTorch Conference North America 2026: Accelerating Open Intelligence Together

PyTorch Conference North America 2026 ran October 20–21 in San Jose with more than 3,500 attendees, organized around the theme of accelerating open intelligence. PyTorch Foundation Executive Director Mark Collier opened both days, framing the event around the Foundation's goal of providing open source software to train and serve models on any chip, in any cloud, for any agent. A Day 0 of training, workshops and co-located events preceded the main program on October 19.
The headline engineering news was TorchTPU, an open source collaboration between Google and Meta that adds a native 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. Migrating a workload can in some cases require only a device change; reaching peak performance may additionally require adjusting shapes, sharding and operations. This is a meaningful signal: torch.compile and eager support mean TPU users get the same programming model as GPU users rather than a separate framework, though the caveat about shapes and sharding suggests the last mile of performance tuning stays hardware-specific.
A second backend, TorchNeuron, targets AWS Trainium and the corresponding Amazon EC2 instances Trn2 and Trn3. Announced as a public beta under Apache 2.0 on October 20 at the conference, it is positioned so that running PyTorch workloads needs a single configuration change. It offers eager execution, torch.compile, and native support for PyTorch APIs including DTensor and torch.distributed, plus libraries such as TorchTitan, spanning research experimentation through production deployment.
Compiler work extended to IBM and AMD hardware. IBM's "From KTIR to SpyreCode" session explained 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 emit SpyreCode. A FlyDSL update on October 21 described an optional backend for AMD MI350-series GPUs that routes eligible operations to optimized kernels through existing PyTorch APIs while keeping existing implementations for workloads it does not support, a fallback pattern that limits risk when adopting a new backend.
TorchTitan, the reference stack for large-scale training, expanded model support across its 5D parallelism stack and added reinforcement learning through TitanRL. TitanRL keeps one model implementation across pretraining, post-training and execution inside vLLM, with asynchronous rollouts and multiple forms of parallelism for both training and generation. The update also covered performance improvements, FlexShard, graph-based training, distributed communication and checkpointing. In practice, a single implementation spanning training and inference matters because it removes a common source of drift between the model a team trains and the one it serves.
Helion's integration with NVIDIA CompileIQ brings compiler controls into kernel autotuning. Developers can evaluate Advanced Control Files alongside kernel configurations such as tile sizes and warp counts, tune offline, and pin the chosen configuration, with compiler defaults still remaining a candidate in the search. ExecuTorch gained a Cortex-M backend, giving a direct path from PyTorch training to on-device inference on Arm Cortex-M CPUs with Arm CMSIS-NN acceleration, supporting all three capability tiers (Scalar, DSP and MVE/Helium) from a single workflow. That matters for embedded and microcontroller targets where memory is measured in kilobytes.
Zephon's TorchTitan integration addresses determinism under changing hardware counts: identical global batches are preserved when training resumes on a different number of GPUs. In a 1B-parameter model test, training resumed from a 16-GPU checkpoint on 8, 32 and 64 GPUs, and the SHA-256 hash of every global batch matched the original run. Zephon tokenizes, packs, shuffles and mixes data inside the training job, avoiding a preprocessed copy of the corpus. My reading is that this targets a real pain point: elastic training usually perturbs data order, which makes comparing runs difficult and complicates reproducibility claims.
Nine organizations joined the Foundation. Nebius, Nscale and Shopify joined as Platinum members; LightSeek Foundation and Lablup as Silver; and SKOPOS INDIA FOUNDATION, Siksha 'O' Anusandhan, Messiah University and Valencia College as Associates. Nebius has supported vLLM with compute for development, testing and benchmarking, including multi-node work on large models such as DeepSeek R1. Nscale arrived via Ray and, following its planned acquisition of Anyscale, the company behind Ray, has committed to continued investment in Ray's development and contributor community, building on Ray's governance under the Foundation since the project joined in October 2025. Collier said open systems learn faster when more people, projects and organizations work together, and that open source ecosystems let breakthroughs travel between organizations rather than staying locked inside one company. Nick Jones, VP of Cloud Native Platform Engineering at Nscale, described treating the framework and the data center as one problem rather than two.
Shopify's Cody Mazza-Anthony presented a continual learning loop using PyTorch and vLLM that compresses production failures into model weights daily, which the company says beats frontier-model quality and cuts serving costs by 96 percent. That is a striking cost figure, though the claim rests on Shopify's own case study and the definition of "frontier-model quality" on its tasks is not specified.
The program featured more than 100 sessions and 144 posters spanning core framework work, training, inference, kernel engineering, on-device deployment and responsible AI. Keynotes included Alban Desmaison and Edward Yang of Meta on PyTorch updates, Chris Lattner of Qualcomm on MAX and Mojo, Ion Stoica of UC Berkeley and Anyscale on Ray, Robert Hundt of Amazon on Trainium's path to native PyTorch, Simon Mo of Inferact on vLLM, and Bill Jia of Google Cloud with Zongwei Zhou of Meta on native PyTorch for Cloud TPUs, among others. Day 0 included PyTorch Associate Training aligned with the PyTorch Certified Associate exam and an Academia to Adoption workshop hosted by the Foundation's OSPO & Academic Outreach Working Group. The 2026 Contributor Awards, sponsored by Red Hat, recognized winners across categories including Outstanding Ambassador (Jiho Kim), Superhero (Thomas Ortner) and Ecosystem Champion (Guangye Yu). Diamond sponsors were AMD, AWS, Crusoe, Google, IBM, Meta, NVIDIA and Qualcomm. The conference returns to San Jose on October 6–7, 2027.
Why it matters: Developers who have kept training on NVIDIA GPUs mainly for ecosystem reasons now have first-party paths to TPUs and Trainium from the same PyTorch code, with performance tuning still required for peak results. Teams running elastic or multi-cloud training should watch the determinism work around Zephon and the RL unification in TitanRL, since both reduce the gap between experimentation and production serving. Infrastructure vendors joining as Platinum members, particularly Nscale's commitment to Ray, suggest the orchestration layer will keep consolidating under Foundation governance, which is my inference from the stated commitments rather than a reported outcome.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
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... The post PyTorch Conference North America 2026: Accelerating Open Intelligence Together appeared first on PyTorch .