NVIDIA CUDA
- Supports CUDA acceleration and TensorRT-oriented optimization paths.
- Best throughput for YOLO-heavy workflows.
Apple Silicon (MPS)
- Uses PyTorch MPS backend where available.
- Good convenience/performance for local labeling and medium workloads.
AMD ROCm
- Requires system ROCm installation before Python packages are effective.
- Use
verify_rocm.py to validate runtime support.
CPU-Only
- Supported for all workflows.
- Prefer background subtraction mode and conservative preview settings.
Portable Defaults
- Start with
YOLO_DEVICE=auto.
- Keep
ENABLE_TENSORRT=false unless CUDA environment is verified.
- Use smaller resize factors for constrained devices.