PyTorch 2.11 Released
6 hours ago
- #PyTorch
- #GPU Optimization
- #Machine Learning
- PyTorch 2.11 release includes 2723 commits from 432 contributors.
- Key features: Differentiable Collectives for Distributed Training, FlexAttention with FlashAttention-4 backend, MPS operator expansion, RNN/LSTM GPU export support, XPU Graph for Intel GPUs.
- MPS improvements include error reporting and expanded operator coverage.
- RNN/LSTM modules can now be exported on GPUs, supporting dynamic shapes.
- ROCm device-side assertions and TopK optimizations enhance AMD GPU performance.
- XPUGraph reduces CPU overhead for Intel GPU workloads.
- FP16 half-precision GEMM support via OpenBLAS for CPU-based deployments.
- CUDA 13 is now the default version; Torchscript is deprecated in favor of torch.export.
- Release cadence increased to every 2 months starting 2026.