rfdetr-ghost: A drop-in performance optimization layer for RF-DETR

rfdetr-ghost is a drop-in performance optimization layer for RF-DETR.

The goal is simple: make RF-DETR significantly faster without requiring users to change their code or learn new APIs.

Users replace:

from rfdetr import ...

with:

from rfdetr_ghost import ...

Everything else stays the same — models, APIs, training/inference code, and workflow.

Under the hood, Ghost patches and replaces performance-critical paths with custom fused CUDA/Triton kernels and other GPU-level optimizations, targeting things like kernel launch overhead, memory movement, and inefficient intermediate operations.

Core idea

One import change. Same API. Faster execution.

The project targets up to ~5× speedup (80% less execution time) on supported workloads, with actual gains depending on the GPU, model, batch size, and operation.

The deeper goal is to demonstrate that significant low-level GPU optimization can be packaged behind a familiar high-level API — allowing existing RF-DETR users to benefit from optimized kernels out of the box, with zero migration effort.

Project positioning

RF-DETR: Python API → GPU

RF-DETR Ghost: Same Python API → Ghost optimization layer → custom fused kernels → GPU

Ghost is inspired by the idea of being invisible to the application but highly visible in the benchmark.