# Ultralytics YOLO gains official AMD GPU support via ROCm and MIGraphX

> **Key takeaway:** Ultralytics YOLO now officially supports AMD GPUs through a ROCm-based backend, with MIGraphX inference delivering up to 67% lower latency than PyTorch for exported ONNX models.

**Published:** 2026-10-26T20:03:12.032130+00:00 
**Source:** AMD ROCm Official Blog 
**Category:** ai-local-edge 
**Canonical URL:** https://fosswire.org/news/ultralytics-yolo-gains-official-amd-gpu-support-via-rocm-and.html 

## Summary
AMD and Ultralytics have completed an official integration that enables the popular YOLO computer vision framework to train, run inference, and deploy on AMD GPUs using ROCm. The collaboration, merged via AMD pull request #24137, exposes AMD hardware through PyTorch's ROCm backend and adds MIGraphX as an optimized inference path for exported ONNX models. Supported tasks span all seven YOLO26 capabilities: object detection, instance and semantic segmentation, classification, pose estimation, oriented bounding boxes, depth estimation, and multi-object tracking.

The integration has been validated across AMD's full GPU stack from data center to embedded: Instinct MI350X, Radeon AI PRO R9700, and Ryzen AI processors including the Max+ PRO 395 and Embedded X100/P185 series. Training is available on Instinct and Radeon discrete GPUs, while inference and MIGraphX deployment cover all tested platforms. Linux x86_64 is required; Windows and WSL are unsupported. A preconfigured Docker image and pip-installable PyTorch wheels with bundled ROCm libraries lower the barrier to entry for developers.

## ROCm integration architecture and deployment paths
- **PyTorch ROCm backend:** AMD GPUs are exposed as `cuda` devices in PyTorch on ROCm, allowing existing Ultralytics code using `device=0` to work unchanged across GPU vendors.
- **MIGraphX inference optimization:** Exported ONNX models can run through ONNX Runtime with AMD's MIGraphX graph engine, achieving up to 67% lower latency than PyTorch inference at equivalent accuracy in Ultralytics validation tests.
- **GPU architecture auto-detection:** The pip install path uses `/sys/class/kfd/kfd/topology/nodes` to detect the GPU's `gfx_target_version` and select matching PyTorch wheels, eliminating manual architecture specification.
- **Self-contained ROCm libraries:** PyTorch wheels bundle required ROCm libraries for the target GPU, so no system-wide ROCm installation is needed.
- **CI validation:** Ultralytics runs AMD GPU tests in its daily continuous integration, ensuring ongoing compatibility as both codebases evolve.

## Why it matters
This integration removes a significant friction point for AMD GPU adoption in the computer vision and edge AI markets, where YOLO dominates real-time object detection workloads. For Linux developers and system builders, it means workstation and server hardware from AMD can now run the most popular vision training pipeline without CUDA dependency, while the MIGraphX path offers a vendor-native optimization route for production deployment on Instinct, Radeon, and Ryzen AI hardware.