mudler/rfdetr-cpp-medium

RF-DETR Medium — GGUF for rfdetr.cpp

GGUF-format weights of Roboflow RF-DETR Medium (detection variant) for use with rfdetr.cpp, a C++/ggml implementation that matches the upstream PyTorch model on CPU.

This repo contains all four standard quantizations of this variant. F16 is the recommended default — same accuracy as F32, 1.85× smaller, and typically the fastest on modern CPUs thanks to ggml's F32×F16 matmul fast path.

Available files

File	Quant	Size (MB)	Recall @ IoU 0.5	Recall @ IoU 0.95	Mean |Δscore|	Latency (median ms, T=8)
rfdetr-medium-f32.gguf	F32	125.0	0.9389	0.8937	0.0065	139.9
rfdetr-medium-f16.gguf ← recommended	F16	67.2	0.9389	0.9032	0.0067	143.9
rfdetr-medium-q8_0.gguf	Q8_0	40.2	0.9389	0.9032	0.0083	145.8
rfdetr-medium-q4_K.gguf	Q4_K	32.5	0.8992	0.7714	0.0199	155.8

All accuracy numbers are computed against the upstream PyTorch reference (rfdetr 1.7.0) on 7 COCO val2017 images at threshold 0.5. Latency is measured with rfdetr-cli bench (8 iters + 3 warmup) at T=8 threads on a single AMD Ryzen 9 9950X3D image (coco_kitchen.jpg, 640x427).

Architecture

• Backbone: DINOv2-small
• Input resolution: 576×576
• Patch size: 14
• Decoder layers: 4
• Object queries: 300
• Task: object detection (boxes only)

Quantization notes

• F32 — full-precision reference, ~120 MB. Bit-exact PyTorch parity.
• F16 — matmul-multiplicand weights only; LayerNorms, conv kernels, embeddings, biases, and layer-scale gammas stay F32. Lossless on this model and consistently the fastest variant on CPU.
• Q8_0 — best size/accuracy tradeoff under F16; ~3× smaller than F32 with effectively identical detections.
• Q4_K — smallest practical quant. Rows with ne[0] % 256 != 0 (the decoder's 128-dim MLP halves, 60 tensors) silently fall back to Q8_0 per ggml's quantizer logic — net compression is still ~3.8× over F32. Use only when the size budget is tight; expect a measurable Recall@0.95 drop relative to F16/Q8_0 (see file table above).

Usage

# 1. Clone + build rfdetr.cpp
git clone https://github.com/mudler/rf-detr.cpp
cd rt-detr.cpp
cmake -B build -DRFDETR_BUILD_CLI=ON && cmake --build build -j

# 2. Download a quant (F16 recommended)
hf download mudler/rfdetr-cpp-medium rfdetr-medium-f16.gguf --local-dir models/

# 3. Run detection
build/bin/rfdetr-cli detect \
    --model models/rfdetr-medium-f16.gguf \
    --input my_image.jpg \
    --threshold 0.5 --threads 8 \
    --output detections.json

Accuracy methodology

All accuracy metrics are computed against the upstream PyTorch reference (rfdetr 1.7.0) on 7 COCO val2017 images at threshold 0.5. Each detection match uses greedy Hungarian-style assignment by IoU (≥ 0.5 lenient, ≥ 0.95 strict) with class equality required.

See BENCHMARK.md and benchmarks/results/accuracy_sweep.json for the full sweep across all 32 (variant × quant) cells.

License

Apache-2.0 — matches the upstream rfdetr license.