GPU-D is a GPU-accelerated implementation of the inverse ray-shooting technique used to generate cosmological microlensing magnification maps. These maps approximate the source plane magnification patterns created by an ensemble of stellar-mass compact objects within a foreground macrolens galaxy. Unlike other implementations, GPU-D solves the gravitational lens equation without any approximation. Due to the high computational intensity and high degree of parallelization inherent in the algorithm, it is ideal for brute-force implementation on GPUs. GPU-D uses CUDA for GPU acceleration and require NVIDIA devices to run.

Herculens models imaging data of strong gravitational lenses. The package supports various degrees of model complexity, ranging from standard smooth analytical profiles to pixelated models and machine learning approaches. In particular, it implements multiscale pixelated models regularized with sparsity constraints and wavelet decomposition, for modeling both the source light distribution and the lens potential. The code is fully differentiable - based on JAX (ascl:2111.002) - which enables fast convergence to the solution, access to the parameters covariance matrix, efficient exploration of the parameter space including the sampling of posterior distributions using variational inference or Hamiltonian Monte-Carlo methods.