Model request: RF-DETR

[cdeil]

Would be great if you could add https://github.com/roboflow/rf-detr .

It's a model that can run on CoreML but with a few issues (see https://github.com/landchenxuan/rf-detr-to-coreml).

Might be a good test case to see if CoreAI can handle it.

[john-rocky]

RF-DETR (nano / small / medium / large, rfdetr 1.7.1) is now running on Core AI — converted, verified and benchmarked in the community zoo:

• Ready-to-run .aimodel bundles: https://huggingface.co/mlboydaisuke/RF-DETR-CoreAI (fp32, Apache-2.0)
• Conversion script + model card: https://github.com/john-rocky/coreai-model-zoo/blob/main/zoo/rf-detr.md
• Measured: 8.6–19.1 ms/frame on an M4 Max GPU (nano→large) and live camera detection on iPhone 17 Pro at ~36 FPS (nano) / ~17 FPS (medium); detections gate set-exact against the PyTorch fp32 reference on CPU and GPU. DETR family, so no NMS anywhere — host post-processing is one sigmoid + top-k.

It was indeed a good stress test. grid_sample has no lowering, but rfdetr 1.7.1 already ships a gather-based bilinear fallback, and the rest of the architecture (windowed DINOv2, two-stage top-k + gather, iterative box refinement) converts cleanly. What actually needed work was four platform bugs, each now filed with a minimal repro: apple/coreai-torch#8 (float-arg arange aborts the converter), apple/coreai-torch#9 (float→int→float cast folding drops truncation), apple/coreai-torch#10 (GPU delegate runs floor/trunc/ceil as identity), apple/coreai-torch#11 (int64-comparison mask chains clobber unrelated live buffers). Numerically-identical workarounds for all four are in the conversion script, with the write-up in https://github.com/john-rocky/coreai-model-zoo/blob/main/knowledge/conversion-guide.md#detection-transformers

[cdeil]

@john-rocky - wow. thank you!