TensorMesh is a fast, differentiable, GPU-native finite element library for PyTorch. It is JIT-free, easy to debug, and end-to-end differentiable, so finite element solvers can sit directly inside a training loop.