Metadata-Version: 2.4
Name: rfdetr-ghost
Version: 0.0.1
Summary: A drop-in performance optimization layer for RF-DETR: same API, custom fused CUDA/Triton kernels underneath.
Keywords: rf-detr,detr,object-detection,cuda,triton,gpu,kernels,inference-optimization
Author: deependujha
Author-email: deependujha <deependujha21@gmail.com>
License-Expression: Apache-2.0
License-File: LICENSE
Classifier: Development Status :: 3 - Alpha
Classifier: Environment :: GPU :: NVIDIA CUDA
Classifier: Intended Audience :: Developers
Classifier: Intended Audience :: Science/Research
Classifier: Operating System :: OS Independent
Classifier: Programming Language :: Python :: 3
Classifier: Programming Language :: Python :: 3.10
Classifier: Programming Language :: Python :: 3.11
Classifier: Programming Language :: Python :: 3.12
Classifier: Programming Language :: Python :: 3.13
Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
Classifier: Topic :: Scientific/Engineering :: Image Recognition
Classifier: Typing :: Typed
Requires-Python: >=3.10
Project-URL: Homepage, https://github.com/deependujha/rfdetr-ghost
Project-URL: Repository, https://github.com/deependujha/rfdetr-ghost
Project-URL: Issues, https://github.com/deependujha/rfdetr-ghost/issues
Description-Content-Type: text/markdown

# `rfdetr-ghost`

**rfdetr-ghost** is a drop-in performance optimization layer for [RF-DETR](https://github.com/roboflow/rf-detr).

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

Users replace:

```python
from rfdetr import ...
```

with:

```python
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.**
