AI BriefingPyTorch Foundation Announces Updates on 6 Major Projects
·2026.07.24 19:01
Key point
The PyTorch Foundation revealed the latest achievements and roadmaps for 6 projects, including PyTorch and vLLM.
Details
In April 2025, the PyTorch Foundation expanded into a multi-project Foundation to drive innovation across the entire AI lifecycle, and it now hosts 6 projects: PyTorch, vLLM, DeepSpeed, Ray, Helion, and Safetensors.
Key PyTorch Updates:
- PyTorch 2.13 Release: Introduced FlexAttention, up to 12x faster on Apple Silicon, added nn.LinearCrossEntropyLoss which cuts memory usage by up to 4x, and support for Python 3.15.
- Distributed Training and Platform: Added a torchcomms backend for large-scale cluster training and FSDP2 communication overlap, and expanded support for ROCm/Arm/Intel XPU.
- On-device AI: Expanded on-device LLM capabilities via ExecuTorch and collaboration with Hugging Face.
Key vLLM Updates:
- Performance Improvements: Significantly improved GPTQ performance through the Model Runner V2 redesign, and established a biweekly release cadence.
- Model Support: Continued Day-zero support for the latest SOTA models such as Kimi K3, Minimax M3, and Qwen 3.8.
- Roadmap: Announced plans to optimize agentic workloads and redesign KV cache behavior in the Q3 2026 roadmap.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Optimizing Gemma 4 Serving with vLLM on Amazon EKS: Cold Start Reduced from 428s to 226s
ZeroModels Released, Compatible with JAX, PyTorch, and TensorFlow
vLLM Publishes Performance Benchmarks for 5 Speculative Decoding Methods on AMD GPUs