[Confirmed working] DeepStream 9.0 + YOLO26 on dGPU — request for official support

[IJLee0812]

DeepStream 9.0 is not yet listed in the README requirements (latest is 8.0). I confirmed this repo's YOLO26 path runs successfully on DS 9.0, and would like to request official support.

Related: #693

Environment (dGPU)

• NVIDIA DeepStream SDK 9.0.0
• Host: Ubuntu 22.04.4 LTS, x86_64
• GPU: NVIDIA GeForce RTX 4090
• NVIDIA Driver 580.95.05
• Container (where DS 9.0 runs): Ubuntu 24.04.3 LTS
  • CUDA 13.1
  • TensorRT 10.14.1.48 (+cuda13.0)

Result

• nvdsinfer_custom_impl_Yolo builds cleanly against DS 9.0 headers/libs — no source changes needed on the parser side; just rebuilt libnvdsinfer_custom_impl_Yolo.so against the DS 9.0 sysroot.
• TensorRT engine generated successfully from the exported YOLO26 ONNX.
• Inference runs end-to-end through DeepStream 9.0; bounding boxes and classes match the DS 8.0 / 7.1 baseline on the same clip.
• No runtime errors, no parser-side API breakage observed on DS 9.0.

ONNX export

Used utils/export_yolo26.py from this repo, with one small local patch required to get the export working on torch >= 2.5:

torch.onnx.export(
    ...,
    dynamo=False,  # Ultralytics models fail with dynamo exporter on torch>=2.5
)

Without dynamo=False, torch.onnx.export picks the new dynamo exporter by default on recent torch versions and fails inside Ultralytics' custom Detect / graph hooks. Explicitly pinning the legacy TorchScript exporter fixes it. Flagging here in case it's useful to fold into the upstream export script.

Notes / caveats

• Tested only on dGPU (RTX 4090, x86_64) — I do not have a Jetson device, so I cannot verify DeepStream 9.0 / JetPack 7.x on Jetson. It would be great if other users could confirm the Jetson path and report back here.
• Only YOLO26 tested on DS 9.0 in my setup; other model families in this repo (RT-DETR / D-FINE / YOLOv8–v13 / YOLO-NAS / etc.) are not yet re-validated on 9.0.

Request

Could you consider:

1. Adding a "DeepStream 9.0 on x86 platform" section to the README requirements list.
2. Bumping the repo description to include 9.0.

Happy to help by sharing the exact build command, the config_infer_primary_yolo26.txt used, or running additional models on DS 9.0 and reporting back. Thanks for maintaining this repo — it's been the reference path for every DeepStream + YOLO integration I've worked on.