PyTorch Foundation: Six Open-Source Projects Accelerate AI Development, FlexAttention Now 12x Faster

PyTorch Foundation brings together six open-source projects

The PyTorch Foundation, a multi-project foundation established in April 2025, today brings together six open-source projects: PyTorch, vLLM, DeepSpeed, Ray, Helion, and other components of the machine learning ecosystem. Shared governance enables coordinated development of tools spanning training, serving, and model optimization.

How much faster is FlexAttention in PyTorch 2.13?

The new PyTorch 2.13 release brings FlexAttention, an attention mechanism that is now about 12x faster on Apple Silicon than in previous versions. The new nn.LinearCrossEntropyLoss function reduces required GPU memory by up to 4x, and the second quarter recorded 4,415 commits, significantly more than comparable earlier periods in the project.

Progress at vLLM, DeepSpeed, Ray Data, and Helion

vLLM has established a stable two-week release cycle and announced its first dedicated conference, running August 24-26 at the Ray Summit. DeepSpeed shipped six releases in the first quarter, and its SuperOffload received an honorable mention for best paper at the ASPLOS 2026 conference. Ray and vLLM on AMD MI325X GPUs achieve a 67 percent savings in prefill-decode disaggregation, while the upcoming Ray Data 2.57 announces a completely new high-performance engine. Helion delivers 10x faster autotuning and state-of-the-art results on NVIDIA Blackwell and Google TPU architectures.