diff --git a/models/common.py b/models/common.py
index 82b348ae..a76fc628 100644
--- a/models/common.py
+++ b/models/common.py
@@ -314,9 +314,11 @@ class DetectMultiBackend(nn.Module):
             net = cv2.dnn.readNetFromONNX(w)
         elif onnx:  # ONNX Runtime
             LOGGER.info(f'Loading {w} for ONNX Runtime inference...')
-            check_requirements(('onnx', 'onnxruntime-gpu' if torch.has_cuda else 'onnxruntime'))
+            cuda = torch.cuda.is_available()
+            check_requirements(('onnx', 'onnxruntime-gpu' if cuda else 'onnxruntime'))
             import onnxruntime
-            session = onnxruntime.InferenceSession(w, None)
+            providers = ['CUDAExecutionProvider', 'CPUExecutionProvider'] if cuda else ['CPUExecutionProvider']
+            session = onnxruntime.InferenceSession(w, providers=providers)
         else:  # TensorFlow model (TFLite, pb, saved_model)
             import tensorflow as tf
             if pb:  # https://www.tensorflow.org/guide/migrate#a_graphpb_or_graphpbtxt