GPU Backends

Backend Detection

hydra_suite.utils.gpu_utils centralizes runtime capability checks for:

• CUDA (CuPy/PyTorch)
• MPS (PyTorch)
• ROCm (PyTorch/CuPy variants where configured)

Where GPU Is Used

• YOLO inference paths (hydra_suite.core.detectors.engine)
• Background operations in supported acceleration paths
• Batch sizing heuristics (hydra_suite.utils.batch_optimizer)

Operational Notes

• Keep device selection explicit in reproducibility-critical runs.
• Use auto for general user workflows.
• TensorRT is optional and environment-sensitive.