diff --git a/.github/workflows/tests.yml b/.github/workflows/tests.yml
new file mode 100644
index 0000000..6078d16
--- /dev/null
+++ b/.github/workflows/tests.yml
@@ -0,0 +1,22 @@
+name: Component tests
+
+on:
+  pull_request:
+  push:
+    branches: [main]
+
+permissions:
+  contents: read
+
+jobs:
+  tests:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+        with:
+          persist-credentials: false
+      - uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0
+        with:
+          python-version: '3.11'
+      - run: python -m pip install pytest
+      - run: python -m pytest -q tests
diff --git a/GraphLib/device.py b/GraphLib/device.py
new file mode 100644
index 0000000..6950265
--- /dev/null
+++ b/GraphLib/device.py
@@ -0,0 +1,14 @@
+"""Tensor-attribute transfer for the legacy DGL graph interface."""
+
+
+def send_graph_to_device(g, device):
+    """Move node/edge attributes in place without removing them before transfer.
+
+    This preserves the legacy helper's graph identity and topology behavior.
+    If a tensor transfer fails, its source attribute remains available.
+    """
+    for name in g.node_attr_schemes():
+        g.ndata[name] = g.ndata[name].to(device, non_blocking=True)
+    for name in g.edge_attr_schemes():
+        g.edata[name] = g.edata[name].to(device, non_blocking=True)
+    return g
diff --git a/GraphLib/dglutil.py b/GraphLib/dglutil.py
index a9102ce..7b6d037 100644
--- a/GraphLib/dglutil.py
+++ b/GraphLib/dglutil.py
@@ -1,3 +1,8 @@
+if __package__:
+    from .device import send_graph_to_device
+else:  # Legacy scripts import dglutil from GraphLib directly.
+    from device import send_graph_to_device
+
 import dgl
 import dgl.function as fn
 import torch as th
@@ -70,15 +75,3 @@ def merge_graphs(graphs, feature_arrays, label_arrays, training_masks):
     labels   = np.concatenate(label_arrays)
     training_masks = np.concatenate(training_masks)
     return g, features, labels, training_masks
-
-def send_graph_to_device(g, device):
-    # nodes
-    labels = g.node_attr_schemes()
-    for l in labels.keys():
-        g.ndata[l] = g.ndata.pop(l).to(DEVICE, non_blocking=True)
-    
-    # edges
-    labels = g.edge_attr_schemes()
-    for l in labels.keys():
-        g.edata[l] = g.edata.pop(l).to(DEVICE, non_blocking=True)
-    return g
diff --git a/README.md b/README.md
index 33779f9..5e0f57a 100644
--- a/README.md
+++ b/README.md
@@ -20,6 +20,19 @@ Start with the [model code](GraphLib/model.py) to inspect the architecture or th
 - [FastAFGCN](https://github.com/lmlearning/FastAFGCN): quantized ONNX inference.
 - [AFSubsample](https://github.com/lmlearning/AFSubsample): framework subsampling and analysis.
 
-## License
+## Component tests
+
+```bash
+python -m pip install pytest
+python -m pytest tests
+```
+
+The device-transfer helper is isolated in `GraphLib/device.py` and remains available
+through `GraphLib.dglutil`. It uses the requested device and keeps the original
+attribute if its transfer fails. The graph is modified in place; earlier successful
+transfers are not rolled back. These tests use graph/tensor protocol doubles and do
+not require DGL or validate GPU execution or model training.
+
+## License
 
 See [LICENSE](LICENSE).
diff --git a/tests/test_device.py b/tests/test_device.py
new file mode 100644
index 0000000..3939681
--- /dev/null
+++ b/tests/test_device.py
@@ -0,0 +1,53 @@
+import pytest
+
+from GraphLib.device import send_graph_to_device
+
+
+class Tensor:
+    def __init__(self, fail=False):
+        self.fail = fail
+        self.calls = []
+
+    def to(self, device, **kwargs):
+        self.calls.append((device, kwargs))
+        if self.fail:
+            raise RuntimeError("transfer failed")
+        return ("converted", device)
+
+
+class Graph:
+    def __init__(self, nodes=None, edges=None):
+        self.ndata = nodes or {}
+        self.edata = edges or {}
+
+    def node_attr_schemes(self):
+        return dict.fromkeys(self.ndata)
+
+    def edge_attr_schemes(self):
+        return dict.fromkeys(self.edata)
+
+
+@pytest.mark.parametrize("device", ["cpu", "cuda:1"])
+def test_transfers_both_attribute_kinds_to_requested_device(device):
+    node, edge = Tensor(), Tensor()
+    graph = Graph({"features": node}, {"weights": edge})
+    assert send_graph_to_device(graph, device) is graph
+    assert graph.ndata == {"features": ("converted", device)}
+    assert graph.edata == {"weights": ("converted", device)}
+    assert node.calls == edge.calls == [(device, {"non_blocking": True})]
+
+
+@pytest.mark.parametrize("attribute_kind", ["ndata", "edata"])
+def test_failed_transfer_keeps_original_attribute(attribute_kind):
+    tensor = Tensor(fail=True)
+    graph = Graph()
+    getattr(graph, attribute_kind)["features"] = tensor
+    with pytest.raises(RuntimeError, match="transfer failed"):
+        send_graph_to_device(graph, "cpu")
+    assert getattr(graph, attribute_kind)["features"] is tensor
+
+
+def test_empty_graph_is_unchanged():
+    graph = Graph()
+    assert send_graph_to_device(g=graph, device="cpu") is graph
+    assert graph.ndata == graph.edata == {}