fix(rocm): use PyTorch+HIP for inference instead of ONNX

Ultralytics' ONNX loader only supports CUDAExecutionProvider (NVIDIA).
On ROCm, it falls back to CPU even though ROCMExecutionProvider is
available. PyTorch + HIP runs natively on AMD GPUs via device='cuda'.

- Change ROCm BackendSpec: onnx → pytorch (skip ONNX export entirely)
- Set YOLO_AUTOINSTALL=0 in detect.py to prevent ultralytics from
  auto-installing onnxruntime-gpu (NVIDIA) at runtime

Author: solderzzc

skills/detection/yolo-detection-2026/scripts/detect.py

@@ -15,12 +15,16 @@
 """
 
 import sys
+import os
 import json
 import argparse
 import signal
 import time
 from pathlib import Path
 
+# Prevent ultralytics from auto-installing packages (e.g. onnxruntime-gpu on ROCm)
+os.environ.setdefault("YOLO_AUTOINSTALL", "0")
+
 # Import env_config — try multiple locations:
 # 1. Same directory as detect.py (bundled copy)
 # 2. DeepCamera repo: skills/lib/

skills/detection/yolo-detection-2026/scripts/env_config.py

@@ -51,9 +51,9 @@ class BackendSpec:
     ),
     "rocm": BackendSpec(
         name="rocm",
-        export_format="onnx",
-        model_suffix=".onnx",
-        half=False,  # ONNX Runtime ROCm handles precision internally
+        export_format="pytorch",     # PyTorch + HIP — ultralytics ONNX doesn't support ROCMExecutionProvider
+        model_suffix=".pt",
+        half=False,
     ),
     "mps": BackendSpec(
         name="mps",

skills/lib/env_config.py

@@ -51,9 +51,9 @@ class BackendSpec:
     ),
     "rocm": BackendSpec(
         name="rocm",
-        export_format="onnx",
-        model_suffix=".onnx",
-        half=False,  # ONNX Runtime ROCm handles precision internally
+        export_format="pytorch",     # PyTorch + HIP — ultralytics ONNX doesn't support ROCMExecutionProvider
+        model_suffix=".pt",
+        half=False,
     ),
     "mps": BackendSpec(
         name="mps",