[GraphBolt][CUDA] Refine pinning tests (#6994)

tests/python/pytorch/graphbolt/impl/test_fused_csc_sampling_graph.py

@@ -1623,6 +1623,8 @@ def test_sample_neighbors_homo(labor, is_pinned):
     0   1   0   0   1
     1   0   0   0   1
     """
+    if F._default_context_str == "cpu" and is_pinned:
+        pytest.skip("Pinning is not meaningful without a GPU.")
     # Initialize data.
     total_num_edges = 12
     indptr = torch.LongTensor([0, 3, 5, 7, 9, 12])
@@ -1631,12 +1633,9 @@ def test_sample_neighbors_homo(labor, is_pinned):
     assert indptr[-1] == len(indices)
 
     # Construct FusedCSCSamplingGraph.
-    graph = gb.fused_csc_sampling_graph(indptr, indices)
-    if F._default_context_str == "gpu":
-        if is_pinned:
-            graph.pin_memory_()
-        else:
-            graph = graph.to(F.ctx())
+    graph = gb.fused_csc_sampling_graph(indptr, indices).to(
+        "pinned" if is_pinned else F.ctx()
+    )
 
     # Generate subgraph via sample neighbors.
     nodes = torch.LongTensor([1, 3, 4]).to(F.ctx())
@@ -1883,6 +1882,8 @@ def test_sample_neighbors_return_eids_homo(labor, is_pinned):
     0   1   0   0   1
     1   0   0   0   1
     """
+    if F._default_context_str == "cpu" and is_pinned:
+        pytest.skip("Pinning is not meaningful without a GPU.")
     # Initialize data.
     total_num_edges = 12
     indptr = torch.LongTensor([0, 3, 5, 7, 9, 12])
@@ -1896,12 +1897,7 @@ def test_sample_neighbors_return_eids_homo(labor, is_pinned):
     # Construct FusedCSCSamplingGraph.
     graph = gb.fused_csc_sampling_graph(
         indptr, indices, edge_attributes=edge_attributes
-    )
-    if F._default_context_str == "gpu":
-        if is_pinned:
-            graph.pin_memory_()
-        else:
-            graph = graph.to(F.ctx())
+    ).to("pinned" if is_pinned else F.ctx())
 
     # Generate subgraph via sample neighbors.
     nodes = torch.LongTensor([1, 3, 4]).to(F.ctx())

tests/python/pytorch/graphbolt/impl/test_torch_based_feature_store.py

@@ -207,7 +207,8 @@ def test_feature_store_to_device(device):
 )
 @pytest.mark.parametrize("idtype", [torch.int32, torch.int64])
 @pytest.mark.parametrize("shape", [(2, 1), (2, 3), (2, 2, 2), (137, 13, 3)])
-def test_torch_based_pinned_feature(dtype, idtype, shape):
+@pytest.mark.parametrize("in_place", [False, True])
+def test_torch_based_pinned_feature(dtype, idtype, shape, in_place):
     if dtype == torch.complex128:
         tensor = torch.complex(
             torch.randint(0, 13, shape, dtype=torch.float64),
@@ -219,10 +220,13 @@ def test_torch_based_pinned_feature(dtype, idtype, shape):
     test_tensor_cuda = test_tensor.cuda()
 
     feature = gb.TorchBasedFeature(tensor)
+    if in_place:
         feature.pin_memory_()
 
         # Check if pinning is truly in-place.
         assert feature._tensor.data_ptr() == tensor.data_ptr()
+    else:
+        feature = feature.to("pinned")
 
     # Test read entire pinned feature, the result should be on cuda.
     assert torch.equal(feature.read(), test_tensor_cuda)