Make sure that benchmark_harness is set before running (pytorch#145532)

Running torch compile with these options causes an error, because the benchmark code isn't generated but is still called.
```
options={'profile_bandwidth_output': 'foo', 'benchmark_harness': False}
```

Pull Request resolved: pytorch#145532
Approved by: https://github.com/eellison

Author: exclamaforte

test/inductor/test_aot_inductor.py

@@ -2089,6 +2089,28 @@ def forward(self, x):
         x = torch.randn(5, device=self.device)
         self.check_model(Model(self.device), (x,))
 
+    def test_profile_benchmark_harness(self):
+        batch_size = 32
+        seq_length = 50
+        hidden_size = 768
+
+        def create_test_fn():
+            def test_fn():
+                inp = torch.randn(
+                    batch_size, seq_length, hidden_size, device=self.device
+                )
+                weight = torch.randn(hidden_size, hidden_size, device=self.device)
+                matmul_output = inp @ weight
+                torch.nn.LayerNorm(hidden_size, device=self.device)(matmul_output)
+                return True
+
+            return test_fn
+
+        fn = torch.compile(
+            options={"profile_bandwidth_output": "foo", "benchmark_harness": False}
+        )(create_test_fn())
+        fn()
+
     def test_with_profiler(self):
         class Model(torch.nn.Module):
             def __init__(self) -> None:

torch/_inductor/graph.py

@@ -2044,7 +2044,7 @@ def _compile_to_module(self) -> ModuleType:
         self.cache_path = path
         self.cache_linemap = linemap  # type: ignore[assignment]
 
-        if config.profile_bandwidth_output:
+        if config.benchmark_harness and config.profile_bandwidth_output:
             # run the inputs code gen to get the bandwidth info
             mod.benchmark_compiled_module(times=1, repeat=1)
         # Logged twice as per https://github.com/pytorch/pytorch/pull/99038#discussion_r1167826029