PyTorch 2.13 is here, with 3,328 commits from 526 contributors and updates across FlexAttention, CuTeDSL, nn.LinearCrossEntropyLoss, torchcomms, FSDP2, Python 3.15 wheels, ROCm, Arm, and XPU. The release blog and notes cover FlexAttention on Apple Silicon with up to ~12x speedup over SDPA on sparse patterns, a deterministic backward path on CUDA, the CuTeDSL "Native DSL" backend for Inductor, nn.LinearCrossEntropyLoss to reduce peak GPU memory by up to 4x, torchcomms for large-cluster training, and FSDP2 communication overlap improvements. On July 22 at 11 a.m. PT, join @albanDesmaison (@Meta), Andrey Talman (@Meta), Piotr Bialecki (@NVIDIA), and Chris Gottbrath for a live 2.13 Q&A. 🔗 Read the release blog, and register for the live Q&A:

PyTorch Foundation supported the ExecuTorch Hackathon in San Francisco, where more than 100 participants across 20+ teams built real-time AI applications using PyTorch and ExecuTorch. Teams built on Snapdragon-powered Samsung Electronics Galaxy S25 Ultra devices, focusing on latency, offline capability, privacy-sensitive processing, energy efficiency, and real-time user experience. Congratulations to the winning teams: 1st Place: SafeScreen AI, an on-device visual safety layer 2nd Place: SixthSense, an assistive wearable that converts visual information into directional haptic signals 3rd Place: Toddle AI, a privacy-first prototype for analyzing toddler walking patterns locally The winning projects showed how local execution can support applications that require immediate feedback, limited connectivity, or sensitive data processing. Read the full recap from @matthew_d_white (PyTorch Foundation), Andrew Caples (@Meta), and Lauren Lunde (@Qualcomm):

PyTorch Foundation is a Gold Sponsor of Agentic AI Summit 2026. Matt White, CTO of PyTorch Foundation, will lead “The Open Agentic Stack,” on building AI systems with open source, open standards, and composability. 🔗

PyTorch-native NeMo AutoModel handles transformer pretraining in @nvidia's end-to-end workflow for building a transaction foundation model. The workflow combines GPU-accelerated data processing and tokenization, decoder-only model pretraining, embedding extraction, and XGBoost fraud classification. On the synthetic @IBM TabFormer dataset, combining raw features with learned embeddings increased Average Precision by 41.76% over the raw-feature baseline. 🔗 Read the full post:

New on the PyTorch Foundation blog: @AMD and @Meta contributors share how PyTorch Monarch was brought to AMD Instinct GPUs with ROCm to support fault tolerant distributed training at scale. The post walks through the ROCm port of Monarch’s GPU runtime and distributed communication stack, then shows how Monarch, TorchFT, and TorchTitan enable healthy replicas to continue training while failed nodes recover and rejoin without a full checkpoint restart. Validation includes Llama 3 8B training on a 128 GPU AMD Instinct MI300 SLURM cluster and a 256 GPU AMD Instinct MI355 Kubernetes cluster. Read the full technical deep dive:

Built on PyTorch, Ray, SGLang, and NVIDIA Megatron-LM, Miles is an open source framework from RadixArk for large-scale LLM reinforcement learning post-training. Miles uses PyTorch for models, numerics, profiling, and extensibility; Ray for orchestration; SGLang for rollout generation; and Megatron-LM for distributed training. The framework supports asynchronous rollout and training, NCCL/RDMA weight synchronization, MoE-aware rollout/training alignment, low-precision recipes, LoRA, fault tolerance, observability, and extension points for custom algorithms and model architectures. 🔗 Read more in our latest blog from the Miles Team:

Joseph Gabriel Lagonsin (ITBrief) reports on Shopify’s Platinum membership in PyTorch Foundation, including Shopify’s plans to contribute upstream engineering expertise and share experience from running machine learning systems in retail and commerce settings. "AI is becoming the operating layer for commerce, and we're convinced that layer needs to be open to reach global scale," said Mikhail Parakhin, Chief Technical Officer, Shopify. "PyTorch is central to how Shopify builds AI today. Joining the Foundation lets us invest in that base directly and help shape it for the agentic era, rather than just building on top of it." The coverage also details Shopify’s use of PyTorch across Sidekick, search and recommendation tools, fraud protection, and foundation model work tied to merchant tools. "We are excited to welcome Shopify to the PyTorch Foundation as our newest Platinum Member," said Mark Collier, Executive Director, PyTorch Foundation. "Shopify operates where AI meets real buyers and sellers every day, and that vantage point is exactly what the Foundation needs as agents become a front door to commerce. Read more:

We’re excited to welcome Mooncake to the PyTorch Ecosystem! Mooncake is designed to solve the “memory wall” in LLM serving. By integrating Mooncake’s high performance KVCache transfer and storage capabilities with PyTorch native inference engines like SGLang, vLLM, and TensorRT-LLM, it unlocks new levels of throughput and scalability for large language model deployments. Mooncake enables prefill decode disaggregation, global KVCache reuse, elastic expert parallelism, and serves as a fault tolerant PyTorch distributed backend. 🔗 #PyTorch# #OpenSourceAI# #LLM# #AIInfrastructure#

🚀 Your next big #OpenSource# conversation starts in Shanghai. Join #KubeCon# + #CloudNativeCon# + #OpenInfraSummit# + #PyTorchCon# China, September 7-9, for three days of technical sessions, community collaboration, & the ideas shaping the #AI# era. ⏰ Register by July 28 to save ¥710 RMB: