marearts-tinydeform: deformable attention in one CUDA call
Open source · CUDA · PyTorch
marearts-tinydeform is our first open-source CUDA kernel. It runs the value projection and multi-scale deformable attention of a DETR-style decoder (RT-DETR, D-FINE, DINO, Deformable-DETR) as one call, gives the same result as the usual PyTorch code, and is up to 4.7× faster than the original Deformable-DETR CUDA op when there are few queries. Apache-2.0, free for any use.