Kareem

Flint turns a PyTorch function into an optimized GPU kernel with just one decorator. Built on Meta’s KernelAgent, Flint uses Spark Open and Wonder to generate forward and backward kernels, check outputs and gradients against the original Python, and benchmark the researcher’s training command. On a two-state recurrent model, Flint made the full training run 2.62x faster than default `torch.compile` on an RTX 4090. The goal was a highly developer-friendly, nearly invisible way to achieve the performance of optimized kernels, without having to write them yourself.