FlashKAN Accelerates KANs with Truncated Power Form, Improves Performance.

Kolmogorov-Arnold Networks (KANs) are a promising alternative to traditional neural networks, using learnable B-spline activations on network edges. However, their primary bottleneck has been the slow Cox-de Boor recursion, which accounts for over 90% of the forward-pass time when evaluating these B-spline activations. FlashKAN addresses this by leveraging the classical truncated power form, which expresses uniform cubic B-splines as a sum of five shifted terms. This allows for a torch.compile-fused implementation that collapses all recursive, span lookup, and scatter-gather operations into a single, highly efficient GPU kernel. The project also introduces bounded-coordinate stabilization to prevent catastrophic cancellation, a historical reason for using the Cox-de Boor recursion. FlashKAN is released as a production-ready, open-source package, serving as a direct drop-in replacement for existing KAN layers, making KANs much more practical for real-world applications.