PyTorch

Transformers v5 going all-in on PyTorch, new ecosystem projects, and major moments from NeurIPS 2025 headline the December PyTorch newsletter. Catch up on the latest from the PyTorch community, including ecosystem updates, PyTorch Conference 2025 recaps, and recent member and community blogs. 🔗 Read the full newsletter: https://hubs.la/Q03YG5_q0 📨 Subscribe to receive the PyTorch newsletter directly in your inbox: https://hubs.la/Q03YG6Jc0 #PyTorch #AIInfrastructure #OpenSourceAI

In this PyTorch Foundation Spotlight, Simon Mo shares why vLLM became a PyTorch Foundation project in May 2025 and how the project has been built on PyTorch from the beginning. He discusses vLLM’s role within a broader ecosystem of model builders and hardware providers, its focus on ease of use and efficiency, and how the project helps users get value quickly while continuing to improve inference efficiency and cost. 🎥 Watch the Spotlight: https://lnkd.in/gUNDgi5z #PyTorch #vLLM #AIInfrastructure #OpenSourceAI

Arm’s Kieran Hejmadi has published a new PyTorch blog introducing a hands-on tutorial series on advanced hardware–software co-design techniques. These approaches help developers train models tailored for resource-constrained edge devices. When deployed on Arm, they gain automatic acceleration through KleidiAI with runtimes like ExecuTorch. What’s inside: • Co-design methods that balance accuracy and model size • Quantization-aware training and extreme quantization • Mixture-of-Experts techniques that cut latency while preserving reasoning 👉 Blog: https://lnkd.in/gdHNQecz Co-produced with Arm Education, Oliver Grainge (University of Southampton), and Prof. Constantine Caramanis (UT Austin). #PyTorch #OpenSourceAI #ExecuTorch #Arm #Quantization #HardwareSoftwareCodesign

NVIDIA Blackwell introduces Cluster Launch Control to enable dynamic scheduling and hardware-driven load balancing. Our latest blog shows how to write TLX kernels using CLC and introduces TLX as the first DSL with dedicated support for it. 🖇️ https://lnkd.in/g_4AvsMJ ✍ Daohang Shi, Hongtao Yu, and Manman Ren (Meta) #PyTorch #GPU #AIInfrastructure #OpenSourceAI

The PyTorch Ecosystem continues to expand with projects that strengthen performance, scalability, and developer workflows across training, inference, AI infrastructure, and scientific machine learning. Our ecosystem update introduces new additions including FlagGems, Kubeflow Trainer, LMCache, DeepInverse, Feast, NeuralOperator, PINA, and verl. It also features sc2bench, a Python package built on PyTorch and torchdistill that supports research on supervised compression for split computing. 🔗 https://hubs.la/Q03Yqllt0 If you’re developing a project for the PyTorch community, you can apply for inclusion in the PyTorch Ecosystem. Read the blog to learn more. #PyTorch #OpenSource #AIInfrastructure #MachineLearning

Ray is built with long-term sustainability in mind, with a focus on neutral governance, openness, and growing alongside its community. In this PyTorch Foundation Spotlight, Richard Liaw shares why Ray joined PyTorch Foundation. Richard points to the ecosystem that has consolidated around PyTorch Foundation, including PyTorch, DeepSpeed, and vLLM, and reflects on what Ray becoming part of the Foundation signals about its role in the AI infrastructure stack and the trust behind it. 🎥 Watch the Spotlight on YouTube: https://lnkd.in/gkpgB3-t #Ray #AIInfrastructure #OpenSourceAI #AI #PyTorch

We had a strong presence at NeurIPS 2025, contributing across the week with workshops, talks, posters, a co-hosted community reception, and ongoing engagement with researchers, engineers, and maintainers from across the AI community. Our recap highlights PyTorch-led workshops on agentic development, on-device and edge AI, a technical session on modular agents and open agentic protocols, poster presentations on Regulatable ML and LLM evaluation workshops, and the Open Source AI Reception co-hosted with Cloud Native Computing Foundation (CNCF) and supported by Anyscale, Featherless AI, Hugging Face, and Unsloth AI. We spent the week connecting with the community around PyTorch, vLLM, DeepSpeed, Ray, and other ecosystem projects, discussing practical deployment, performance optimization, and what’s ahead. Read our NeurIPS 2025 recap: https://hubs.la/Q03YmkN30 Workshops & Technical Sessions: Joseph Spisak (Meta), Davide Testuggine (Meta), Aksel Joonas Reedi (Hugging Face), Zach Wentz (Meta), Daniel Han (Unsloth AI), Sanyam Bhutani (Meta), Cagatay Bilgin (Meta), Matthias Lecher, Andrey Tovchigrechko (Meta), Varun Khare (NimbleEdge), Vijay Janapa Reddi (Harvard University), Karan Goel (Cartesia), Mitesh Patel, Ph.D. (NVIDIA) #PyTorch #NeurIPS2025 #OpenSourceAI #AIInfrastructure

NeuralOperator is now part of the PyTorch Ecosystem, bringing a dedicated PyTorch library for learning neural operators and enabling applications of AI to science and engineering problems that involve mappings between function spaces. Built on PyTorch, NeuralOperator translates neural operator theory into a practical, open source library for scientific computing, with support for state-of-the-art operator learning models that integrate directly into existing PyTorch workflows. 🔗 Read more: https://lnkd.in/giGuNYEv 👥 Anima Anandkumar, Jean Kossaifi, Valentin Duruisseaux, and David Pitt. #PyTorch #AIInfrastructure #OpenSourceAI

DeepSpeed.ai serves as a research library for advancing ideas in ML training. In this PyTorch Foundation Spotlight, Zhipeng(Jason) Wang, PhD, (DeepSpeed) explains why joining the PyTorch Foundation is beneficial for the project. Zhipeng discusses PyTorch as the default and most popular ML framework across the AI community, and how being part of this ecosystem supports new feature development and a broader range of AI training workloads for researchers and developers. 🎥 Watch the Spotlight: https://lnkd.in/gvfbswbU #PyTorch #DeepSpeed #OpenSourceAI #AIInfrastructure