The PyTorch Foundation published a July 22, 2026 update on Q2 progress across its six hosted projects. Highlights include PyTorch 2.13’s FlexAttention delivering up to 12x faster performance on Apple Silicon, vLLM’s stable biweekly release cadence, DeepSpeed’s new Ulysses parallelism integration, Ray’s work scaling frontier model training, Helion’s cross-hardware attention kernels with 10x faster autotuning, and Safetensors adding GIL-free serialization with Python 3.14 support.
