diff --git a/src/sagemaker/local/image.py b/src/sagemaker/local/image.py
index 38d963bfda..1fbc93fca0 100644
--- a/src/sagemaker/local/image.py
+++ b/src/sagemaker/local/image.py
@@ -759,7 +759,9 @@ def _create_docker_host(self, host, environment, optml_subdirs, command, volumes
         # for GPU support pass in nvidia as the runtime, this is equivalent
         # to setting --runtime=nvidia in the docker commandline.
         if self.instance_type == "local_gpu":
-            host_config["runtime"] = "nvidia"
+            host_config["deploy"] = {
+                "resources": {"reservations": {"devices": [{"capabilities": ["gpu"]}]}}
+            }
 
         if command == "serve":
             serving_port = (
diff --git a/tests/unit/test_image.py b/tests/unit/test_image.py
index 020f648834..e7bde99610 100644
--- a/tests/unit/test_image.py
+++ b/tests/unit/test_image.py
@@ -574,8 +574,10 @@ def test_container_has_gpu_support(tmpdir, sagemaker_session):
     )
 
     docker_host = sagemaker_container._create_docker_host("host-1", {}, set(), "train", [])
-    assert "runtime" in docker_host
-    assert docker_host["runtime"] == "nvidia"
+    assert "deploy" in docker_host
+    assert docker_host["deploy"] == {
+        "resources": {"reservations": {"devices": [{"capabilities": ["gpu"]}]}}
+    }
 
 
 def test_container_does_not_enable_nvidia_docker_for_cpu_containers(sagemaker_session):