marearts-tinydeform: deformable attention in one CUDA call

marearts-tinydeform: deformable attention in one CUDA call

marearts-tinydeform is our open-source CUDA kernel for DETR-style decoders: value projection and multi-scale deformable attention in one call, same result as the PyTorch code. With few queries it is up to 4.7x faster than the original Deformable-DETR CUDA op (RT-DETR, D-FINE, DINO).

Read the full article for free

First visit? No sign-up needed. We'll email you a 6-digit sign-in code. No password required.

Continue with email