PyTorch

On Oct 19 at PyTorch Conference North America 2026 in San Jose, attendees spent a full day in our instructor-led PyTorch Associate Training preparing for the PyTorch Certified Associate (PTCA) exam. Check out what the attendees had to say here: https://bit.ly/4Bvhb2j #PyTorchCon

TorchNeuron is now open source. Our new @pytorch.org backend translates code into operations executable on AWS Trainium with minimal or no code changes. Learn how the team used AI agents and hundreds of thousands of models to validate it.

Faster streaming speech recognition on @Arm edge devices: see how ExecuTorch, SME2, FP16 and QC8W balance accuracy and latency—delivering up to 2.12× faster processing on Pixel 11. Read our latest blog: https://bit.ly/4zcnjLi

DeepSpeed virtual office hours on October 27 cover features and benchmark results presented at PyTorch Conference and the proposed Q4 roadmap. 9 AM PT / 12 PM ET https://zoom-lfx.platform.linuxfoundation.org/meeting/93902569995?password=7d9c4fc9-3efa-4715-88f0-df8a6deb008b

3,500+ attendees came together to accelerate open intelligence at #PyTorchCon 2026. Catch up on new members, TorchTPU, TorchNeuron, TorchTitan, and community highlights: https://pytorch.org/blog/pytorch-conference-north-america-2026-accelerating-open-intelligence-together/

PyTorch Foundation welcomes its 2026 cohort of 24 ambassadors, creating more opportunities for people to discover open source AI, develop practical skills, and contribute. Meet the ambassadors: pytorch.org/blog/meet-th...

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. Learn more: pytorch.org/blog/torchti...

Today's Demo Theater at #PyTorchCon, Community Expo: 13 demos from AWS, Baseten, Verda, d-Matrix, Clockwork, Intel, Huawei, Lightning AI, Inferact, Lemurian Labs, Rebellions, HPE, and Anyscale. https://events.linuxfoundation.org/pytorch-conference-north-america/program/schedule/?topic=demo-theater

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: https://bit.ly/3W1zwUv

Unwind at the Flare Party & Poster Presentations tonight at #PyTorchCon, 6–7:30 PM, Community Expo. Meet poster presenters, enjoy good food (sliders, biryani, churros, etc.), and join Pub Trivia at 6:30. https://events.linuxfoundation.org/pytorch-conference-north-america/program/schedule-posters/

Demo Theater this afternoon at #PyTorchCon, Community Expo: 2:00 IBM, Spyre 2:15 Datadog, Ray observability 2:30 Modal, Street Fighter III 3:55 Arm, ExecuTorch 4:10 Red Hat, quantization https://events.linuxfoundation.org/pytorch-conference-north-america/program/schedule/?topic=demo-theater

Stop by the PyTorch Foundation booth in the Community Expo and tell us what you're building. Ask about PyTorch, vLLM, DeepSpeed, Ray, Helion, or Safetensors, or share what you're seeing across the open source AI stack. #PyTorchCon

How does a PyTorch workload become something an AI accelerator can actually execute? In our latest blog, the IBM Spyre team explores the open source compiler technology connecting PyTorch to IBM's Spyre accelerator. Read more: https://bit.ly/4jvcTSf

Google & Meta launch TorchTPU for native PyTorch support on Google TPUs! 🔥True eager execution & torch.compile backend ⚡Fast inference via vLLM & SGLang 🏋️Near-linear scaling for Llama 3 405B & Qwen 3 MoE via TorchTitan Read more here: https://bit.ly/4hV19Ho

We're excited to welcome Nebius as a Platinum Member of the PyTorch Foundation. Nebius brings production-scale experience in AI infrastructure, inference, and post-training, and has long supported vLLM with compute for development and benchmarking. Read the announcement: https://bit.ly/4BfxuAl

#PyTorchCon North America 2026 starts today in San Jose! Can't be here in person? Stream the keynotes live: https://events.linuxfoundation.org/pytorch-conference-north-america/program/keynote-livestream/ Talks and workshops across PyTorch, vLLM, DeepSpeed, Ray, Helion, and Safetensors.

#PyTorchCon starts tomorrow in San Jose to accelerate open intelligence. Explore the program: https://events.linuxfoundation.org/pytorch-conference-north-america/program/schedule/ Join us: https://events.linuxfoundation.org/pytorch-conference-north-america/register/

Squeeze extra performance out of GPU kernels that are already autotuned to the limit by diving below the kernel level into compiler controls. NVIDIA CompileIQ and PyTorch Helion unite two distinct optimization layers into a single joint search space - learn more here: https://bit.ly/4i8Oxew

More hardware choice for PyTorch developers: TorchNeuron registers @amazonwebservices.bsky.social Trainium as a native device, built on the same extension points the community created to make PyTorch hardware-agnostic. Open sourcing at PyTorch Conference on Oct. 20. https://bit.ly/4i3P9Ud

Ravi Gupta, Shiksha Patel, and Rishi Madduri (AMD) will present “Efficient Inference of Masked Diffusion Language Models on AMD XDNA2 NPU” at #PyTorchCon North America 2026. https://hubs.la/Q04v4SL60

Looking for sessions to catch at #PyTorchCon North America 2026? Chris Hoge (NVIDIA), co-chair of the PyTorch Ecosystem Working Group, highlights five talks for developers working on open source training and performance. https://hubs.la/Q04v4SL60

Join us at PyTorch Conference North America 2026, where Zesheng Zong (Huawei), PyTorch Foundation Ambassador and Accelerator Integration Working Group Co-Chair, will share key technical updates on backend integration. Join us in San Jose: https://hubs.la/Q04v4SL60 #PyTorchCon

ExecuTorch is now an official Zephyr external module. Zephyr Project’s walkthrough covers setup and runs quantized MobileNetV2 on an Alif Ensemble E8 DevKit. https://www.zephyrproject.org/executorch-integrated-with-zephyr-rtos/

Zephon’s TorchTitan integration preserved identical global batches when a 1B-parameter model resumed from a 16-GPU checkpoint on 8, 32, and 64 GPUs. Read more: https://pytorch.org/blog/bringing-elastically-deterministic-data-loading-to-torchtitan-with-zephon/

Ray All the Way Down: A Heterogeneous, Elastic PyTorch Training Stack at LinkedIn Join Tommy Li and Tao Huang from LinkedIn at PyTorch Conference North America 2026 for their lightning talk, Ray All the Way Down, on October 20. Register now: https://hubs.la/Q04v4SL60 #PyTorchCon

Today, Meta is open sourcing the weights for Muse Spark Open, the most capable US open-weights model to date. It’s integrated with the PyTorch ecosystem: it ships with Helion kernels, and TorchTitan can be used to further tune the model. More in the blog: https://pytorch.org/blog/muse-spark-open/

Bringing ExecuTorch to Arm Cortex-M CPUs In our latest blog, Arm and Meta walk through how they have brought ExecuTorch to Arm Cortex-M CPUs. Learn more about this technical milestone here: https://bit.ly/4zu39Mz

PyTorch Conference North America 2026 program committee member Ricardo Aravena (Nunchux AI) calls #PyTorchCon a place to connect with people building the future of machine learning and learn practical ideas from the community. Register: https://hubs.la/Q04v4SL60

In this blog, IBM shares how they extended PyTorch Inductor for the IBM Spyre Accelerator, leveraging PyTorch's compiler to address challenges in data movement, memory management, & work distribution while optimizing for Spyre's unique architecture. Learn more here: https://bit.ly/4dCdqxW

Look who it is and he does such a good job explaining why PyTorch conference is the place to be

Running large Mixture-of-Experts models across pods creates two major challenges: expert traffic and KV-cache transfer. At #PyTorchCon, Ravi Gupta, Rishi Madduri, and Shiksha Patel (AMD) will present how MoRI + vLLM tackles both. https://hubs.la/Q04v4SL60

Connect with Madrid’s PyTorch community on October 14, 6:00–8:30 pm, at a free meetup to share real-world experiences and discuss technical questions. https://pytorch.org/series/pytorch-meetups-madrid-spain/

High-Velocity GPU Kernel Authoring with CUTLASS Python At PyTorch Conference North America 2026, Guray Ozen from NVIDIA will present recent improvements to CUTLASS for Python. Join us on October 20-21 in San Jose: https://hubs.la/Q04v4SL60 #PyTorchCon

NVIDIA recsys-examples now provides an end-to-end HSTU inference workflow with Dynamo-Triton. PyTorch AOTI compiles HSTU models to native C++, while FlexKV stores attention states to boost efficiency. On an RTX PRO 6000 Blackwell GPU, this yields up to 5.93x lower latency. https://bit.ly/4AT5Vws

Optimizing kernels can be time-consuming and require deep expert knowledge. KernelAgent is a multi-agent harness that can write and optimize kernels for you. At #PyTorchCon, Kaiming Cheng (Meta) will present hardware-guided GPU kernel optimization. https://hubs.la/Q04v4SL60

Read our latest blog from Nvidia to learn how NVIDIA Dynamo integrates its router, inference engine, and KV-cache manager to deliver seamless end-to-end serving for agentic workloads: https://bit.ly/4emlt21

Bridging the ExecuTorch Adoption Gap Through Developer Learning Join Matt Cossins from Arm at PyTorch Conference North America 2026 in San Jose, CA for a Birds of a Feather discussion session on ExecuTorch. Join us on October 20-21: https://hubs.la/Q04v4SL60 #PyTorchCon

How can a single framework support an increasingly diverse range of AI hardware? In a new PyTorch Foundation blog, contributors from IBM share how Spyre was integrated using existing PyTorch abstractions for devices, streams, memory management, & compilation. Learn more: https://bit.ly/4y7RkKU

Replacing CUDA with FBTriton for Table Batched Embedding (TBE) kernels delivered massive gains for our rec sys models. Meta contributors achieved up to 1.28x faster forward passes, 2x faster backward passes, and a huge boost in developer velocity. Read more: https://bit.ly/4dsVgyC

Think DeepSpeed is only good for ZeRO and its usability, not training throughput? At #PyTorchCon North America 2026, Masahiro Tanaka will cover tensor, sequence, and expert parallelism in DeepSpeed and walk through benchmarks. https://hubs.la/Q04v4SL60

Red Hat, NVIDIA AI, and IBM are hosting “PyTorch, Powering the Enterprise,” a PyTorch Conference Day 0 event on Oct. 19 covering PyTorch, vLLM, Ray, Helion, and llm-d: https://bit.ly/4znEDgl Register for #PyTorchCon NA: https://bit.ly/3U6WZ69

Over the last two years, Meta has consolidated PyTorch's media stack across three libraries: TorchCodec, TorchVision, and TorchAudio. Read the full technical debrief to learn more about what has changed and why: https://bit.ly/3VMBwzN

The PyTorch Accelerator Integration Working Group is working to establish standard hardware enablement mechanisms across the platform to unify cross-backend testing, profiling, compiler support, and distributed execution. Read the full technical recap to learn more: https://bit.ly/4hpHZsY

PyTorch Foundation Ambassador Abdulsalam Bande will present a poster on ERRC (Entropy-Reinvested Residual Correction for Tensor-Parallel LLM Communication) at PyTorch Conference North America 2026. Join us in San Jose, Oct 20-21: https://hubs.la/Q04v88dJ0 #PyTorchCon

Modern AI workloads rely on specialized GPU kernels, and no single backend is optimal across every operator. At #PyTorchCon, Liz Li and Jiahui Cao (AMD) will present how FlyDSL integrates with TorchInductor while preserving the torch.compile experience. https://hubs.la/Q04v4SL60

Helion, a PyTorch-native kernel DSL, powers a vLLM linear backend with per-shape autotuning. On Hopper, tuning plus hybrid dispatch beat vLLM’s default CUTLASS and DeepGEMM backends. https://pytorch.org/blog/building-a-high-performance-and-portable-vllm-linear-backend-with-helion/

NEW! Take the PyTorch Certified Associate (PTCA) Certification Pathway which combines focused learning modules & the PyTorch Certified Associate (PTCA) certification exam together in one structured experience through @lfeducation.bsky.social Learn more: pytorch.org/blog/new-pat...

PyTorch’s latest blog covers Jagged Flash Attention on NVIDIA B200 with TLX. On GEM’s jagged shapes, the TLX kernel is about 13% faster on the forward pass and about 50% faster on the backward pass than the May 2026 version of FA4. https://bit.ly/4dhV84P