pytorch.org

24 articles in the Truth Foundry index. Each links out to the original.

• PyTorch-Triton 3.7 Adds TLX Plugin Extensions for Persistent Kernels · PyTorch-Triton 3.7 enables Meta's TLX plugin for persistent GEMM kernels on NVIDIA and AMD GPUs
• Meta Engineers Fuse Normalization Kernels to Boost LLM Training Speed · Meta researchers fuse normalization operations with GEMM kernels to reduce memory overhead in LLM training.
• PyTorch 2.13 Adds FlexAttention to Apple Silicon and New Distributed Training Features · PyTorch 2.13 introduces FlexAttention on Apple Silicon, CuTeDSL backend, and FSDP2 communication overlap.
• PyTorch's Dynamic Test Infrastructure Explained for Developers · PyTorch's test framework dynamically generates device and dtype-specific tests, complicating local debugging but enabling scalable validation.
• ExecuTorch Hackathon Highlights On-Device AI Innovations · ExecuTorch Hackathon showcases real-time AI applications running locally on Snapdragon-powered devices using PyTorch.
• Shopify Becomes Platinum Member of PyTorch Foundation to Shape AI Future · Shopify joins PyTorch Foundation as Platinum member to influence open-source AI development.
• RadixArk Unveils Miles: A PyTorch-Native Framework for Large-Scale LLM RL Post-Training · RadixArk launches Miles, an open-source framework integrating SGLang, Megatron-LM, and Ray for scalable LLM reinforcement learning.
• PyTorch Launches Cross-Repository CI Relay for Automated Downstream Testing · PyTorch introduces CRCR to automate CI dispatch and results reporting across its ecosystem of downstream repositories.
• TokenSpeed-Kernel Enables Multi-Silicon LLM Inference via Portable APIs · TokenSpeed-Kernel decouples LLM inference runtimes from hardware-specific kernels for AMD and NVIDIA GPUs.
• SGLang Achieves 5x Throughput Boost for DeepSeek-V4 on NVIDIA GB300 · Optimized SGLang stack delivers 5x higher throughput for DeepSeek-V4 on NVIDIA GB300 GPUs
• KubeCon + PyTorch Conference China Schedule Released for September 2026 · Major cloud native and AI conferences in Shanghai announce their September 2026 schedules.
• PyTorch Foundation Opens Nominations for 2026 Contributor Awards · PyTorch Foundation invites community nominations for 2026 awards recognizing diverse contributions to its projects.
• Helion Kernels Boost vLLM FP8 Inference on NVIDIA H100 and B200 GPUs · Helion provides a PyTorch-native workflow for developing fused GPU kernels, delivering performance improvements for vLLM inference with Qwen3 models.
• DeepSpeed Integrates Muon Optimizer for Faster Large Model Training · DeepSpeed adds native support for the Muon Optimizer to improve training efficiency for frontier AI models.
• LinkedIn Re-architects DuaLip Solver with PyTorch for Extreme-Scale Optimization · LinkedIn deployed a GPU-accelerated PyTorch version of its DuaLip linear programming solver to achieve massive speedups.
• PyTorch Compile Achieves 10x Speedup via Kernel Fusion Optimization · PyTorch's Inductor compiler uses kernel fusion to eliminate memory overhead and achieve up to 10x faster model execution.
• TokenSpeed Sets 580 TPS Record with Qwen3.5-397B-A17B on GPUs · TokenSpeed inference engine achieves 580 tokens per second with Qwen3.5-397B-A17B model on GPU clusters.
• Alibaba Cloud Joins PyTorch Foundation as Platinum Member and Governing Board Representative · Alibaba Cloud becomes a Platinum member of the PyTorch Foundation, joining its Governing Board to advance AI infrastructure.
• Facebook Research Unveils TLX Block Attention Kernel for Blackwell GPUs · Facebook Research introduces TLX Block Attention, achieving 2.5x backward speedup on NVIDIA B200 via warp-specialized Triton kernels.
• PyTorch Foundation Expands Ambassador Program with Global Recruitment Push · PyTorch Foundation seeks new ambassadors to support community growth, prioritizing underrepresented regions.
• PyTorch Docathon 2026 Yields 150+ Merged Pull Requests from Community · Over 150 merged pull requests improved PyTorch documentation after the 2026 Docathon.
• PyTorch 2.11 Adds Native aarch64 CUDA Wheels, Ending vLLM Workarounds · PyTorch 2.11 enables direct CUDA wheel installation on aarch64 Linux, simplifying vLLM deployment on NVIDIA GH200 and GB200 systems.
• PyTorch 2.12 Unifies Graph APIs, Boosts CUDA Performance by 100x · PyTorch 2.12 introduces a device-agnostic graph API and significant CUDA performance improvements.
• Arm Labs Demonstrate Efficient Edge AI Inference on CPUs and NPUs with ExecuTorch · Arm provides Jupyter labs showcasing ExecuTorch deployment on Cortex-A, Cortex-M, and Ethos-U NPUs for local AI inference.