[ETHOSN] Only use mock inference when whole graph is offloaded (#12296)

The mock inference functionality is only supported when the whole
graph is offloaded to the NPU, otherwise it can result in undefined
behaviour. This patch makes sure the mock inference functionality is
not run on test cases where some parts of the graph are not offloaded
to the NPU, while ensuring the module is still built as a sanity check.

Change-Id: I27052d118ff976f9adbfc3f5b5b96185318e1573

tests/python/contrib/test_ethosn/test_topologies.py

@@ -22,7 +22,7 @@
 import tvm
 from tvm import relay
 from tvm.testing import requires_ethosn
-from tvm.relay.op.contrib.ethosn import Available
+from tvm.relay.op.contrib.ethosn import Available, ethosn_available
 from . import infrastructure as tei
 
 
@@ -76,19 +76,27 @@ def get_model(input_shape, dtype, var_names):
 
         expected_host_ops = 1 if tei.get_ethosn_api_version() == 2205 else 0
         npu_partitions = 2 if tei.get_ethosn_api_version() == 2205 else 1
-        outputs.append(
-            tei.build_and_run(
-                mod,
-                inputs,
-                1,
-                {},
-                npu=npu,
-                expected_host_ops=expected_host_ops,
-                npu_partitions=npu_partitions,
+
+        # Mock inference is only supported when the whole graph is offloaded to the NPU
+        if tei.get_ethosn_api_version() == 2205 and ethosn_available() == Available.SW_ONLY:
+            tei.build(
+                mod, {}, npu=npu, expected_host_ops=expected_host_ops, npu_partitions=npu_partitions
+            )
+        else:
+            outputs.append(
+                tei.build_and_run(
+                    mod,
+                    inputs,
+                    1,
+                    {},
+                    npu=npu,
+                    expected_host_ops=expected_host_ops,
+                    npu_partitions=npu_partitions,
+                )
             )
-        )
 
-    tei.verify(outputs, dtype, 2)
+    if outputs:
+        tei.verify(outputs, dtype, 2)
 
 
 @requires_ethosn
@@ -118,7 +126,6 @@ def get_model(dtype):
         return out
 
     np.random.seed(0)
-    outputs = []
     inputs = {
         "x": tvm.nd.array(
             np.random.randint(
@@ -128,9 +135,12 @@ def get_model(dtype):
     }
     model = get_model(dtype)
     mod = tei.make_module(model, {})
-    outputs.append(
+
+    # Mock inference is only supported when the whole graph is offloaded to the NPU
+    if ethosn_available() == Available.SW_ONLY:
+        tei.build(mod, {}, npu=True, expected_host_ops=1, npu_partitions=2)
+    else:
         tei.build_and_run(mod, inputs, 1, {}, npu=True, expected_host_ops=1, npu_partitions=2)
-    )
 
 
 @requires_ethosn
@@ -218,19 +228,31 @@ def get_model(shape, dtype, splits, axis):
 
             expected_host_ops = 1 if tei.get_ethosn_api_version() == 2205 else 0
             npu_partitions = 2 if tei.get_ethosn_api_version() == 2205 else 1
-            outputs.append(
-                tei.build_and_run(
+
+            # Mock inference is only supported when the whole graph is offloaded to the NPU
+            if tei.get_ethosn_api_version() == 2205 and ethosn_available() == Available.SW_ONLY:
+                tei.build(
                     mod,
-                    inputs,
-                    2,
                     {},
                     npu=npu,
                     expected_host_ops=expected_host_ops,
                     npu_partitions=npu_partitions,
                 )
-            )
+            else:
+                outputs.append(
+                    tei.build_and_run(
+                        mod,
+                        inputs,
+                        2,
+                        {},
+                        npu=npu,
+                        expected_host_ops=expected_host_ops,
+                        npu_partitions=npu_partitions,
+                    )
+                )
 
-        tei.verify(outputs, dtype, 0)
+        if outputs:
+            tei.verify(outputs, dtype, 0)
 
 
 @pytest.mark.skipif(