Delete the model after checking accuracy in-process

Summary:
When the model is not deleted, there will be problem when running the model and model+cudagraph in the same process.

Note that blue_reels_vdd_v3 does not support CUDA Graph. This is to demonstrate that we can run a model within a process (without accuracy checking)

Reviewed By: davidberard98

Differential Revision: D49688178

fbshipit-source-id: 8594f7a88bba221d1339646f2a6aea7138621bc3

torchbenchmark/util/env_check.py

@@ -359,7 +359,7 @@ def forward_pass(mod, inputs, contexts, _collect_outputs=True):
             return mod(**inputs)
         else:
             return mod(*inputs)
-    
+
 
 def forward_and_backward_pass(mod, inputs, contexts, optimizer, collect_outputs=True):
     cloned_inputs = clone_inputs(inputs)

torchbenchmark/util/experiment/metrics.py

@@ -149,4 +149,6 @@ def get_model_accuracy(model_config: TorchBenchModelConfig, isolated: bool=True)
         return accuracy
     else:
         model = load_model(accuracy_model_config)
-        return model.accuracy
+        accuracy = model.accuracy
+        del model
+        return accuracy