Add GPU-resident YOLO26 video inference example (rocDecode + MIGraphX)

[itikhono]

Summary

New Python example under AI/MIGraphX/gpu_resident_yolo26_pipeline/ demonstrating a zero-copy decode-to-detection pipeline on AMD GPUs:

This is the companion code for the upcoming ROCm blog post Building a GPU-Resident YOLO26 Object Detection Pipeline on the AMD Radeon AI PRO R9700 GPU (ROCm/rocm-blogs#355).

• rocDecode / rocPyDecode decodes the H.264 bitstream on the on-chip VCN engine and exposes the surface via DLPack.
• PyTorch + DLPack wraps it as a zero-copy CUDA tensor and runs YOLO26 letterbox preprocessing on the active HIP stream.
• MIGraphX runs inference of the FP16-compiled YOLO26 graph asynchronously on the same HIP stream and returns detections directly into a pre-allocated PyTorch tensor.
• OpenCV draws the surviving bounding boxes onto each frame and writes the annotated MP4 output.

An OpenCV CPU-decode path (--decoder opencv) is included as a baseline.

Files

README.md, main.py, prepare_model.py, requirements.txt,
.gitignore, data/peloton_sample_ai_gen.mp4 (1080p H.264 sample), images/result_boxes.jpg.

Test plan

Tested inside rocm/pytorch:rocm7.2.2_ubuntu22.04_py3.10_pytorch_release_2.10.0:

• python3 prepare_model.py produces model.mxr
• python3 main.py --decoder rocdecode -> 248 fps end-to-end
• python3 main.py --decoder opencv -> 195 fps end-to-end

tested on AMD Radeon AI PRO R9700 GPU

[itikhono]

@ROCm/rocm-examples-owners could you help with reviewing this sample please?

[zichguan-amd]

Thanks, this LGTM. Is this something that we can/want to add to our CI testing?

[itikhono]

Thanks, this LGTM. Is this something that we can/want to add to our CI testing?

I will double check if we can add a new test