Ultralytics v8.4.72 Released 
Summary
We’ve released Ultralytics v8.4.72, a small but important stability update focused on more reliable TensorRT INT8 exports on certain RTX GPUs 
The headline fix resolves an export crash affecting some format="engine", int8=True workflows, with a few additional improvements for export environment compatibility, CI maintenance, and docs polish. If you export models for deployment, especially with TensorRT INT8, this is a great release to pick up.
New Features and Key Improvements
Fixed TensorRT INT8 export crashes on some RTX GPUs
The most important fix in this release comes from PR #24876 by @glenn-jocher.
On some RTX systems, exporting with format="engine", int8=True could crash when ONNX Runtime exposed conflicting TensorRT execution providers. This release updates the INT8 calibration path to use CUDA with CPU fallback, avoiding the provider combination that triggered the failure.
Why it matters:
- More reliable TensorRT INT8 export on affected NVIDIA RTX hardware
- Fewer unexpected crashes before engine build completes
- Better experience for production-oriented export pipelines
Improved CUDA 12 export image compatibility
In PR #24877, @glenn-jocher pinned onnxruntime-gpu to <1.27.0 in the export Docker image.
Newer ONNX Runtime GPU builds expect CUDA 13, while the current export image still uses CUDA 12.8. This pin helps prevent version mismatch breakage and makes containerized export workflows more predictable.
CI maintenance update
PR #24873 by @UltralyticsAssistant updates GitHub Actions actions/checkout from v6 to v7.
This is a straightforward maintenance improvement for the project automation pipeline.
Documentation Updates
Docs refresh and cleanup
A few documentation improvements also landed in this release:
- PR #24863 by @RizwanMunawar refreshes the Queue Management guide with a newer Ultralytics YOLO26 video and updated wording
- PR #24875 by @onuralpszr simplifies the Triton C++ example README
- PR #24851 by @miles-deans-ultralytics fixes duplicate page title tags for account settings and YOLO config docs
Impact
Who should update?
This release is especially useful if you:
- Export YOLO models to TensorRT INT8
- Use RTX GPUs for deployment workflows
- Rely on Docker-based export environments
If your workflow is mostly training, validation, or standard FP16 export, the changes are smaller, but the overall reliability improvements still make this a worthwhile update.
What’s not changing?
This release does not introduce:
- A new model release
- New training features
- Architecture changes
The focus here is squarely on export stability, environment compatibility, and documentation polish.
Quick Try
To upgrade:
pip install -U ultralytics
If you’re working with export workflows, this is a good time to retest your TensorRT INT8 pipeline with your current models, including Ultralytics Platform deployment and export flows if you use them.
Full Changelog
You can review the complete diff in the full changelog for v8.4.72 or browse the v8.4.72 release page.
Thanks 
Big thanks to everyone who contributed fixes and improvements to this release:
@glenn-jocher, @RizwanMunawar, @onuralpszr, @miles-deans-ultralytics, and @UltralyticsAssistant.
Please give v8.4.72 a try and let us know how it works for your workflow, especially if you’ve been affected by TensorRT INT8 export issues on RTX hardware 