We have an RF-DETR face detection model (exported to ONNX) we’d like to compile to DFP on SDK 2.2. The compiler fails on GridSample, TopK, GatherElements, Where (ops from the multi-scale deformable attention decoder). Given you already have an RT-DETRv2 subgraph handler (architecturally very similar), would a custom NCE extension be possible?
Hello @Miguel,
Thank you for reaching out and working with MX3 product.
A custom Neural Compiler Extension for RF-DETR is technically feasible. However, on the current-generation MX3 platform, support for some transformer-specific operators and graph patterns is still limited, particularly in deformable-attention decoder blocks.
The RT-DETR extension you referenced was developed primarily as an internal engineering prototype to evaluate this class of architecture. It accelerates only a portion of the RT-DETR graph rather than providing complete end-to-end model support, so it is not currently offered as a general-purpose public extension.
We are continuing to expand transformer support as part of our next-generation product work.
For face detection on MX3 today, a YOLO-based detector is likely to be the most straightforward and effective option. In our experience, transformer-based detectors can add significant architectural complexity for this particular use case without necessarily providing a final deployment advantage.
You may find this real-time multi-face recognition pipeline useful as a starting point: