PyTorch Foundation Projects Detail Q2 Updates Across Six Hosted Tools

The PyTorch Foundation now hosts six projects, and each one published a quarterly update on July 22, 2026. The foundation became a multi-project organization in April 2025, and the current lineup is PyTorch, vLLM, DeepSpeed, Ray, Helion, and Safetensors. The updates cover performance work, hardware enablement, and community health across the board.

PyTorch 2.13 and On-Device Expansion

The core PyTorch project logged 4,415 commits in Q2, and the number of open issues is decreasing. The community released PyTorch 2.13 and started the development cycle for 2.14. The 2.13 release brings FlexAttention to Apple Silicon, with the project reporting up to roughly 12 times faster performance than SDPA. A new CuTeDSL Inductor backend targets NVIDIA GPUs, and a memory-saving nn.LinearCrossEntropyLoss reduces peak GPU memory by up to 4 times. Python 3.15 wheel support includes free-threaded builds. On the distributed side, the release adds the torchcomms backend and FSDP2 communication overlap for large-cluster training, plus broader ROCm, Arm, and Intel XPU support. ExecuTorch is expanding on-device LLM capabilities with weekly examples and a partnership with Hugging Face to streamline model ingestion.

vLLM, DeepSpeed, and Ray

vLLM reached a stable bi-weekly release cadence and completed a redesign of Model Runner V2, which produced substantial performance improvements on GPTQ. The project offers day-zero support for models including Kimi K3, Minimax M3, and Qwen 3.8. Its Q3 2026 roadmap is organized around special interest groups, focusing on production agentic workloads and high-interactivity premium tokens. The core engine work includes finishing the Flat Model and Model Runner V2 migrations and redesigning the scheduler and KV cache. Large-scale serving goals target top performance on AgentX, better KV cache offloading, and smarter prefix caching for multi-turn agents. Other tracks cover faster speculative decoding, production-ready KV-cache compression, quicker CI, model performance, vLLM-Omni, and reinforcement learning. The first vLLM Conference will run at Ray Summit in San Francisco from August 24 to 26.

DeepSpeed kept its bi-weekly schedule with six releases in the first quarter and stabilized monthly virtual office hours. The project replaced Intel GPU support (IPEX) with direct Torch.xpu integration. Updates include mixed precision fixes, AMD SDMA optimization, automatic sequence parallelism, and the Muon optimizer. The Ulysses sequence parallelism algorithm was integrated into Hugging Face libraries including Trainer, Accelerate, and TRL. DeepSpeed received a best-paper honorable mention at ASPLOS 2026 for its SuperOffload work and had three research papers accepted, covering automatic sequence parallelism with PyTorch 2.0 compile technology, energy analysis and offloading overheads on Grace Hopper C2C systems, and efficient offloading mechanics for model training.

Ray focused on production hardening for large-scale workloads, including faster actor scheduling, native RDMA support, topology-aware scheduling for rack-aware actor placement, and improved support for GB200s and GB300s. The team also optimized workflows for reinforcement learning, multimodal data, and video processing. Ray was used to build frontier models including MAI-Thinking-1, Composer 2.5, and Nemotron 3 Ultra. Other highlights include inference performance improvements with Ray plus vLLM, better cluster stability via resource isolation, improved data pipeline reliability, and prefill-decode disaggregation with Ray plus vLLM on AMD MI325X. A new high-performance engine for Ray Data is scheduled to debut in version 2.57, and the team is building weight-syncing integrations with vLLM and SkyRL to remove post-training bottlenecks. Ray Summit runs August 24 to 26 in San Francisco.

Helion and Safetensors

Helion introduced a CuTeDSL backend for NVIDIA GPUs and a Pallas backend for TPUs, built in partnership with Google and debuted at Google AI Systems DevLabs. Using the same attention kernel, Helion reports state-of-the-art performance on NVIDIA Blackwell GPU, outperforming FlashAttention-4, and on Google TPU, outperforming the hand-written Attention kernel from the Tokamax library. The project also introduced an LLM-guided autotuner that demonstrates a 10 times improvement in tuning efficiency. Helion held its first hackathon with NVIDIA and GPU Mode and a tutorial at the PLDI conference, and it was announced in General Availability at the PyTorch Conference Europe.

Safetensors rolled out Global Interpreter Lock (GIL)-Free Serialization in June, adding parallel processing for better multi-threaded performance. The project expanded its CI pipeline to cover Python 3.14 and 3.14t. New Metal Performance Shaders (MPS) fast-load paths introduce native Metal buffers within the torch MPS backend to minimize model loading latency. New dtypes offer greater precision and flexibility. Looking ahead, Safetensors is working on CUDA fast-path loading, io_uring, and GPU Direct Storage (GDS) integrations. The initial step is multithreaded pread operations into pinned host buffers to accelerate data transfers. The project is evaluating whether to remove reliance on framework-managed storage like the from_file memory mapping in torch, since the transition depends on DLPack handoffs that may not be universally supported.

The foundation also announced upcoming events: KubeCon plus CloudNativeCon plus OpenInfra Summit plus PyTorch Conference China 2026 in Shanghai from September 7 to 9, and PyTorch Conference North America 2026 in San Jose from October 20 to 21.