Restrict ONNX ExecutionProviders (#18400)

Signed-off-by: Glenn Jocher <glenn.jocher@ultralytics.com>
Signed-off-by: UltralyticsAssistant <web@ultralytics.com>
Co-authored-by: UltralyticsAssistant <web@ultralytics.com>

.github/workflows/ci.yaml

@@ -102,11 +102,11 @@ jobs:
         python-version: ["3.11"]
         model: [yolo11n]
     steps:
-      - uses: astral-sh/setup-uv@v5
       - uses: actions/checkout@v4
       - uses: actions/setup-python@v5
         with:
           python-version: ${{ matrix.python-version }}
+      - uses: astral-sh/setup-uv@v5
       - name: Install requirements
         shell: bash # for Windows compatibility
         run: |

tests/conftest.py

@@ -74,10 +74,10 @@ def pytest_terminal_summary(terminalreporter, exitstatus, config):
 
     # Remove files
     models = [path for x in ["*.onnx", "*.torchscript"] for path in WEIGHTS_DIR.rglob(x)]
-    for file in ["bus.jpg", "yolo11n.onnx", "yolo11n.torchscript"] + models:
+    for file in ["decelera_portrait_min.mov", "bus.jpg", "yolo11n.onnx", "yolo11n.torchscript"] + models:
         Path(file).unlink(missing_ok=True)
 
     # Remove directories
     models = [path for x in ["*.mlpackage", "*_openvino_model"] for path in WEIGHTS_DIR.rglob(x)]
-    for directory in [TMP.parents[1] / ".pytest_cache", TMP] + models:
+    for directory in [WEIGHTS_DIR / "path with spaces", TMP.parents[1] / ".pytest_cache", TMP] + models:
         shutil.rmtree(directory, ignore_errors=True)

tests/test_cli.py

@@ -59,7 +59,8 @@ def test_rtdetr(task="detect", model="yolov8n-rtdetr.yaml", data="coco8.yaml"):
     run(f"yolo train {task} model={model} data={data} --imgsz= 160 epochs =1, cache = disk fraction=0.25")
     run(f"yolo predict {task} model={model} source={ASSETS / 'bus.jpg'} imgsz=160 save save_crop save_txt")
     if TORCH_1_9:
-        run(f"yolo predict {task} model='rtdetr-l.pt' source={ASSETS / 'bus.jpg'} imgsz=160 save save_crop save_txt")
+        weights = WEIGHTS_DIR / "rtdetr-l.pt"
+        run(f"yolo predict {task} model={weights} source={ASSETS / 'bus.jpg'} imgsz=160 save save_crop save_txt")
 
 
 @pytest.mark.skipif(checks.IS_PYTHON_3_12, reason="MobileSAM with CLIP is not supported in Python 3.12")

tests/test_python.py

@@ -576,11 +576,11 @@ def test_model_embeddings():
 @pytest.mark.skipif(checks.IS_PYTHON_3_12, reason="YOLOWorld with CLIP is not supported in Python 3.12")
 def test_yolo_world():
     """Tests YOLO world models with CLIP support, including detection and training scenarios."""
-    model = YOLO("yolov8s-world.pt")  # no YOLO11n-world model yet
+    model = YOLO(WEIGHTS_DIR / "yolov8s-world.pt")  # no YOLO11n-world model yet
     model.set_classes(["tree", "window"])
     model(SOURCE, conf=0.01)
 
-    model = YOLO("yolov8s-worldv2.pt")  # no YOLO11n-world model yet
+    model = YOLO(WEIGHTS_DIR / "yolov8s-worldv2.pt")  # no YOLO11n-world model yet
     # Training from a pretrained model. Eval is included at the final stage of training.
     # Use dota8.yaml which has fewer categories to reduce the inference time of CLIP model
     model.train(

tests/test_solutions.py

@@ -3,18 +3,20 @@
 import cv2
 import pytest
 
+from tests import TMP
 from ultralytics import YOLO, solutions
+from ultralytics.utils import ASSETS_URL, WEIGHTS_DIR
 from ultralytics.utils.downloads import safe_download
 
-MAJOR_SOLUTIONS_DEMO = "https://github.com/ultralytics/assets/releases/download/v0.0.0/solutions_ci_demo.mp4"
+DEMO_VIDEO = "solutions_ci_demo.mp4"
-WORKOUTS_SOLUTION_DEMO = "https://github.com/ultralytics/assets/releases/download/v0.0.0/solution_ci_pose_demo.mp4"
+POSE_VIDEO = "solution_ci_pose_demo.mp4"
 
 
 @pytest.mark.slow
 def test_major_solutions():
     """Test the object counting, heatmap, speed estimation and queue management solution."""
-    safe_download(url=MAJOR_SOLUTIONS_DEMO)
+    safe_download(url=f"{ASSETS_URL}/{DEMO_VIDEO}", dir=TMP)
-    cap = cv2.VideoCapture("solutions_ci_demo.mp4")
+    cap = cv2.VideoCapture(str(TMP / DEMO_VIDEO))
     assert cap.isOpened(), "Error reading video file"
     region_points = [(20, 400), (1080, 400), (1080, 360), (20, 360)]
     counter = solutions.ObjectCounter(region=region_points, model="yolo11n.pt", show=False)  # Test object counter
@@ -42,8 +44,8 @@ def test_major_solutions():
     cap.release()
 
     # Test workouts monitoring
-    safe_download(url=WORKOUTS_SOLUTION_DEMO)
+    safe_download(url=f"{ASSETS_URL}/{POSE_VIDEO}", dir=TMP)
-    cap1 = cv2.VideoCapture("solution_ci_pose_demo.mp4")
+    cap1 = cv2.VideoCapture(str(TMP / POSE_VIDEO))
     assert cap1.isOpened(), "Error reading video file"
     gym = solutions.AIGym(line_width=2, kpts=[5, 11, 13], show=False)
     while cap1.isOpened():
@@ -59,9 +61,9 @@ def test_instance_segmentation():
     """Test the instance segmentation solution."""
     from ultralytics.utils.plotting import Annotator, colors
 
-    model = YOLO("yolo11n-seg.pt")
+    model = YOLO(WEIGHTS_DIR / "yolo11n-seg.pt")
     names = model.names
-    cap = cv2.VideoCapture("solutions_ci_demo.mp4")
+    cap = cv2.VideoCapture(TMP / DEMO_VIDEO)
     assert cap.isOpened(), "Error reading video file"
     while cap.isOpened():
         success, im0 = cap.read()

ultralytics/nn/autobackend.py

@@ -192,14 +192,14 @@ class AutoBackend(nn.Module):
                 check_requirements("numpy==1.23.5")
             import onnxruntime
 
-            providers = onnxruntime.get_available_providers()
+            providers = ["CPUExecutionProvider"]
-            if not cuda and "CUDAExecutionProvider" in providers:
+            if cuda and "CUDAExecutionProvider" in onnxruntime.get_available_providers():
-                providers.remove("CUDAExecutionProvider")
+                providers.insert(0, "CUDAExecutionProvider")
-            elif cuda and "CUDAExecutionProvider" not in providers:
+            elif cuda:  # Only log warning if CUDA was requested but unavailable
-                LOGGER.warning("WARNING ⚠️ Failed to start ONNX Runtime session with CUDA. Falling back to CPU...")
+                LOGGER.warning("WARNING ⚠️ Failed to start ONNX Runtime with CUDA. Using CPU...")
                 device = torch.device("cpu")
                 cuda = False
-            LOGGER.info(f"Preferring ONNX Runtime {providers[0]}")
+            LOGGER.info(f"Using ONNX Runtime {providers[0]}")
             if onnx:
                 session = onnxruntime.InferenceSession(w, providers=providers)
             else: