PyTorch

🚨 Today’s the day! Registration for #PyTorchCon Europe increases €100 TONIGHT after 23:59 CET. Join us 7-8 April in Paris. 🎟 Grab your pass: events.linuxfoundation.org/pytorch-confer…

We’re excited to introduce TorchSpec, a torch-native framework for scalable speculative decoding training developed by the TorchSpec and Mooncake teams. By streaming hidden states from inference engines to training workers via Mooncake, TorchSpec enables fully disaggregated pipelines where inference and training scale independently. 🔗 Read our latest blog from TorchSpec & Mooncake teams: pytorch.org/blog/torchspec… @lightseekorg @KT_Project_AI #PyTorch #TorchSpec #Mooncake #OpenSourceAI

Want to build differentiable computational physics Code leveraging AI? NVIDIA Warp is a framework for accelerated simulation, data generation, and spatial computing that bridges CUDA and Python. Warp enables developers to write high-performance kernels as regular Python functions that are JIT-compiled into efficient code for execution on the GPU. Warp can be easily made differentiable through the Warp native support for automatic differentiation. Straightforward to integrate with optimization or training workflows while remaining interoperable with frameworks like PyTorch, JAX, and NumPy. 🖇️ Read the full post: developer.nvidia.com/blog/build-acc… #PyTorch #OpenSourceAI #AI #Inference #Innovation

@Nvidiadev @matthew_d_white @narendasan @PKUWZP @robertnishihara @_seemethere @mr_purelycloud @alband @malfet @joespeez @ptrblck_de @syedahmedz 1PM: Custom GPU Kernels w/ Triton & TLX w/ Cicie Wang, Hongtao Yu, Manman Ren, Nick Riasanovsky, Shane Nay, Yueming Hao Last chance to connect with the community! 🧡

@Nvidiadev @matthew_d_white @narendasan @PKUWZP @robertnishihara @_seemethere @mr_purelycloud @alband @malfet @joespeez @ptrblck_de @syedahmedz 1PM: Join the Ecosystem w/ Alban Desmaison 2PM: Start Contributing w/ Alban 3PM: CuteDSL CUDA Kernels w/ Vicki Wang, Manas Sahni 4PM: SGLang Inference w/ Baizhou Zhang, Hao Lu 5PM: Monarch Framework w/ Ali Sol, Saurabh Mishra 6PM: TorchComms w/ Davide Italiano, Matthias Reso

Heading to #NVIDIAGTC next week? Let’s talk @PyTorch. 🚀 We’re bringing the community to San Jose. Drop by Booth #338 to meet expert developers and core maintainers in person. Scaling, inference, foundation models, and OSS contributions. Full schedule below 👇 #PyTorch

Great to see @AMD select vLLM as one of the designated inference frameworks for the GPU MODE Hackathon. 🎉 The challenge: push Kimi K2.5 1T FP4 end-to-end inference performance on 8× AMD Instinct MI355X — using vLLM or AMD ATOM. Grand prize: $650,000. What makes this different: winning optimizations must be mergeable into AMD ATOM or vLLM upstream. Improvements that land in vLLM benefit the whole community. Phase 1 (kernel optimization) runs through April 6. More details ⬇️

⏰ Last call! Registration for #PyTorchCon Europe, 7-8 April in Paris, goes up €100 tomorrow at 23:59 CET. 🪧 Don't miss a thing - check out the Poster Presentations: bit.ly/4bPqWNY 🎟 Register now: bit.ly/4bUWj91

We’re excited to share Generalized Dot-Product Attention (GDPA) — a production-driven attention kernel designed specifically for large-scale recommendation systems (RecSys). Proposed in our recent paper, GDPA replaces softmax with a flexible activation tailored for real-world RecSys traffic patterns and has been deployed in Meta’s largest recommendation model, GEM. 🔗 Read our latest blog: pytorch.org/blog/generaliz… By redesigning attention around production characteristics rather than benchmark assumptions, GDPA achieves 2× forward speedup (1,145 BF16 TFLOPs, ~97% tensor core utilization), 1.6× backward speedup, and up to 3.5× forward speedup vs. FA4 under short K/V settings on NVIDIA B200. This work demonstrates how real production traffic can fundamentally reshape kernel design. ✍ Jiaqi Xu, Han Xu, Junqing Zhou, Devashish Shankar, Xiaoyi (Leo) Liu, Shuqi Yang #PyTorch #OpenSourceAI #GDPA #GEM

3️⃣ days left before rates rise! Registration for #PyTorchCon Europe increases €100 after 20 March. 🐦 Join us in Paris, 7-8 April, where you'll find friends with shared interest at BoFs: hubs.la/Q046LJWJ0 🎟 Register: hubs.la/Q046LL3H0

#ExecuTorch addresses fragmented native deployment for #AI agents as a #PyTorch native platform. It enables voice models across CPU, GPU, and NPU on Android, iOS, Linux, macOS & Windows 🔗 pytorch.org/blog/building-…

🚀 Put your brand in front of the #PyTorch community. Sponsor #PyTorchCon Europe, 7–8 April in Paris, and connect with the researchers, engineers & #ML leaders building the next generation of #AI. Showcase your tech. Meet top talent. Build real partnerships. 🤝 Explore opportunities: hubs.la/Q046LpSr0

The #GenAI & Multimodal track at #PyTorchCon Europe explores building & scaling generative & multimodal models using #PyTorch. Learn more: hubs.la/Q046Lxy40 🎟 Register → hubs.la/Q046LfC90

Before we head to Paris for PyTorch Conference EU 2026, we’re looking back on PyTorch Conference 2025 keynotes from visionary AI leaders in our community. We’re starting with Eli Uriegas (@_seemethere) from Meta on testing PyTorch at scale: 11,000 commits and 794 million minutes of CI/CD compute in 2025 alone. This session recognizes the partners working across competitive lines to keep the ecosystem stable for everyone. Watch the full keynote: youtu.be/xWjXsP1E5mQ?si… See you in Paris: pytorch.org/event/pytorch-… #PyTorch #OpenSourceAI #PyTorchConf

🎉 Congrats to @nvidia on the release of Nemotron 3 Super — day-0 support in vLLM v0.17.1! Verified on NVIDIA GPUs. 120B hybrid MoE, only 12B active at inference. Big upgrades over the previous Nemotron Super: - 5x higher throughput - 2x higher accuracy on Artificial Analysis Intelligence Index - Multi-Token Prediction (MTP) for faster long-form generation - Configurable thinking budget — dial accuracy vs token cost per task - 1M token context window Supports BF16, FP8, and NVFP4. Fully open: weights, datasets, recipes. Blog: vllm.ai/blog/nemotron-… 🤝 Thanks @NVIDIAAIDev Nemotron team and vLLM community contributors!

MXFP8 training for MoEs on GB200s enables a 1.3x speedup with equivalent convergence versus BF16. This #PyTorch update via TorchAO and TorchTitan on Crusoe Cloud details gains from dynamically quantized grouped GEMMs. #AI #OpenSource 🔗 pytorch.org/blog/mxfp8-tra… ✍️ @vega_myhre, @_xmfan, @drisspg, Chinmay Baikar

Community matters. ❤️ Join the Women & Non-Binary in #PyTorch Lunch at #PyTorchCon Europe, 7-8 April, in Paris. 🍽️ Learn more: hubs.la/Q046t8KK0 ⏳ Registration goes up €100 on 20 March. 🎟 Get your pass now! hubs.la/Q046tdRT0