← all releases
PyTorch (Meta / Linux Foundation)·May 13, 2026·v2.12·frameworks
PyTorch 2.12
PyTorch 2.12 brings major performance improvements including up to 100x faster batched linear algebra (linalg.eigh) on CUDA via updated cuSOLVER API selection, the new torch.accelerator.Graph API for optimized hardware execution, experimental CUDA 13.2 support, and broad updates across compilation, distributed systems, export, and graph capture.
Deep LearningPyTorchCUDAGPU ComputingFrameworkOpen Source
overview
PyTorch 2.12 brings major performance improvements including up to 100x faster batched linear algebra (linalg.eigh) on CUDA via updated cuSOLVER API selection, the new torch.accelerator.Graph API for optimized hardware execution, experimental CUDA 13.2 support, and broad updates across compilation, distributed systems, export, and graph capture.
key features
- 01Up to 100x faster batched linalg.eigh on CUDA via cuSOLVER backend update
- 02New torch.accelerator.Graph API for optimized hardware execution
- 03Experimental CUDA 13.2 support and updates across compilation and distributed systems
related productsbrowse all →