diff --git a/backend/open_webui/__init__.py b/backend/open_webui/__init__.py
index 30e83b198..743656d9e 100644
--- a/backend/open_webui/__init__.py
+++ b/backend/open_webui/__init__.py
@@ -39,6 +39,18 @@ def serve(
                 "/usr/local/lib/python3.11/site-packages/nvidia/cudnn/lib",
             ]
         )
+        try:
+            import torch
+            assert torch.cuda.is_available(), "CUDA not available"
+            typer.echo("CUDA seems to be working")
+        except Exception as e:
+            typer.echo(
+                "Error when testing CUDA but USE_CUDA_DOCKER is true. "
+                "Resetting USE_CUDA_DOCKER to false and removing "
+                f"LD_LIBRARY_PATH modifications: {e}"
+            )
+            os.environ["USE_CUDA_DOCKER"] = "false"
+            os.environ["LD_LIBRARY_PATH"] = ":".join(LD_LIBRARY_PATH)
     import open_webui.main  # we need set environment variables before importing main
 
     uvicorn.run(open_webui.main.app, host=host, port=port, forwarded_allow_ips="*")
diff --git a/backend/open_webui/env.py b/backend/open_webui/env.py
index d99a80df4..df5597cbf 100644
--- a/backend/open_webui/env.py
+++ b/backend/open_webui/env.py
@@ -36,7 +36,18 @@ except ImportError:
 USE_CUDA = os.environ.get("USE_CUDA_DOCKER", "false")
 
 if USE_CUDA.lower() == "true":
-    DEVICE_TYPE = "cuda"
+    try:
+        import torch
+        assert torch.cuda.is_available(), "CUDA not available"
+        DEVICE_TYPE = "cuda"
+    except Exception as e:
+        cuda_error = (
+            "Error when testing CUDA but USE_CUDA_DOCKER is true. "
+            f"Resetting USE_CUDA_DOCKER to false: {e}"
+        )
+        os.environ["USE_CUDA_DOCKER"] = "false"
+        USE_CUDA = "false"
+        DEVICE_TYPE = "cpu"
 else:
     DEVICE_TYPE = "cpu"
 
@@ -56,6 +67,9 @@ else:
 log = logging.getLogger(__name__)
 log.info(f"GLOBAL_LOG_LEVEL: {GLOBAL_LOG_LEVEL}")
 
+if "cuda_error" in locals():
+    log.exception(cuda_error)
+
 log_sources = [
     "AUDIO",
     "COMFYUI",