add comparison of pcc/atol of all outputs

add _extract_outputs function to ModelTester that subclasses may need to
implement. This function should return a tuple of torch.Tensors

Trim unused code in tests/utils.py

move verify_against_golden to verify.py

replace self.model with self.framework_model and self.compiled_model

tests/models/MobileNetV2/test_MobileNetV2.py

@@ -47,7 +47,7 @@ def test_MobileNetV2(record_property, mode, nightly):
     else:
         cc.compile_depth = CompileDepth.TTNN_IR
 
-    tester = ThisTester(model_name, mode, compiler_config=cc)
+    tester = ThisTester(model_name, mode, compiler_config=cc, required_atol=0.03)
     results = tester.test_model()
     if mode == "eval":
         # Print the top 5 predictions

tests/models/beit/test_beit_image_classification.py

@@ -34,6 +34,9 @@ def append_fake_loss_function(self, outputs):
     def get_results_train(self, model, inputs, outputs):
         return inputs["pixel_values"].grad
 
+    def _extract_outputs(self, output_object):
+        return (output_object.logits,)
+
 
 @pytest.mark.parametrize(
     "mode",
@@ -57,14 +60,18 @@ def test_beit_image_classification(record_property, model_name, mode, nightly):
     else:
         cc.compile_depth = CompileDepth.TTNN_IR
 
-    tester = ThisTester(model_name, mode, compiler_config=cc)
+    atol = 0.05
+    tester = ThisTester(model_name, mode, compiler_config=cc, required_atol=atol)
     results = tester.test_model()
 
     if mode == "eval":
         logits = results.logits
 
         # model predicts one of the 1000 ImageNet classes
         predicted_class_idx = logits.argmax(-1).item()
-        print("Predicted class:", tester.model.config.id2label[predicted_class_idx])
+        print(
+            "Predicted class:",
+            tester.framework_model.config.id2label[predicted_class_idx],
+        )
 
     record_property("torch_ttnn", (tester, results))

tests/models/mgp-str-base/test_mgp_str_base.py

@@ -32,6 +32,9 @@ def _load_inputs(self):
         inputs["pixel_values"] = inputs["pixel_values"].to(torch.bfloat16)
         return inputs
 
+    def _extract_outputs(self, output_object):
+        return output_object.logits
+
 
 @pytest.mark.parametrize(
     "mode",
@@ -52,7 +55,7 @@ def test_mgp_str_base(record_property, mode, nightly):
     else:
         cc.compile_depth = CompileDepth.TTNN_IR
 
-    tester = ThisTester(model_name, mode, compiler_config=cc)
+    tester = ThisTester(model_name, mode, relative_atol=0.02, compiler_config=cc)
     results = tester.test_model()
 
     if mode == "eval":

tests/models/openpose/test_openpose_v2.py

@@ -32,17 +32,19 @@ def get_image_tensor():
     return input_batch
 
 
+# NOTE: If we are to decompose BatchNorm2d (which we must in order to lower to SHLO)
+# we need to run in float32 to maintain precision.
 class ThisTester(ModelTester):
     def _load_model(self):
         # Create PyBuda module from PyTorch model
         model = ptcv_get_model("lwopenpose2d_mobilenet_cmupan_coco", pretrained=True)
-        model = model.to(torch.bfloat16)
+        model = model.to(torch.float32)
         return model
 
     def _load_inputs(self):
         input_batch = [get_image_tensor()]
         batch_input = torch.cat(input_batch, dim=0)
-        batch_input = batch_input.to(torch.bfloat16)
+        batch_input = batch_input.to(torch.float32)
         return batch_input
 
 

tests/models/perceiver_io/test_perceiver_io.py

@@ -34,6 +34,9 @@ def _load_inputs(self):
         }
         return arguments
 
+    def _extract_outputs(self, output_object):
+        return (output_object.logits,)
+
 
 @pytest.mark.parametrize(
     "mode",
@@ -52,7 +55,7 @@ def test_perceiver_io(record_property, mode, nightly):
     else:
         cc.compile_depth = CompileDepth.TTNN_IR
 
-    tester = ThisTester(model_name, mode, compiler_config=cc)
+    tester = ThisTester(model_name, mode, relative_atol=0.02, compiler_config=cc)
     results = tester.test_model()
     if mode == "eval":
         logits = results.logits

tests/models/squeeze_bert/test_squeeze_bert.py

@@ -24,6 +24,9 @@ def _load_inputs(self):
         inputs = self.tokenizer("Hello, my dog is cute", return_tensors="pt")
         return inputs
 
+    def _extract_outputs(self, output_object):
+        return (output_object.logits,)
+
 
 @pytest.mark.parametrize(
     "mode",