AI Edge Torch: High Performance Inference of PyTorch Models on Mobile Devices- Google Developers Blog

Introduces AI Edge Torch, a new tool that enables direct conversion of PyTorch models to TensorFlow Lite (TFLite) runtime for on-device inference. It offers improved CPU performance and initial GPU support compared to existing workflows like ONNX2TF, with validated compatibility on over 70 models from torchvision, timm, torchaudio, and HuggingFace. The tool supports over 70% of core_aten operators and integrates with TorchDynamo and torch.export. It also highlights partnerships with Qualcomm for a new QNN delegate that significantly accelerates inference on their silicon.