Fast-soft-sort provides soft, differentiable versions of sorting and ranking operations that can be integrated into ML training pipelines. The library exposes functions like soft_sort and soft_rank with a regularization_strength parameter that controls how hard or soft the sorting behavior is, enabling end-to-end gradient-based optimization where ranking metrics need to be minimized. It offers native implementations for TensorFlow, JAX, and PyTorch, making it a utility library for researchers working on ranking losses, learning-to-rank models, and neural network architectures that depend on ordering operations.