diff --git a/smdebug/tensorflow/keras.py b/smdebug/tensorflow/keras.py
index 9a02eb0bb..b9855a29b 100644
--- a/smdebug/tensorflow/keras.py
+++ b/smdebug/tensorflow/keras.py
@@ -539,7 +539,11 @@ def _save_layer_input_and_outputs(self):
                 if self._is_collection_being_saved_for_step(CollectionKeys.LAYERS)
                 else set()
             )
-            self._save_tensor_to_file(export_name, tensor.numpy(), input_collection)
+            if hasattr(tensor, "numpy"):
+                self._save_tensor_to_file(export_name, tensor.numpy(), input_collection)
+            else:
+                self.logger.warn("cannot save layer values during forward pass with tf.function")
+                return
             # Save Output
             tensor = self.saved_layers[layer_name].layer_output
             export_name = get_export_name_for_keras(layer_name, tensor_type="output", tensor=tensor)
@@ -549,7 +553,8 @@ def _save_layer_input_and_outputs(self):
                 if self._is_collection_being_saved_for_step(CollectionKeys.LAYERS)
                 else set()
             )
-            self._save_tensor_to_file(export_name, tensor.numpy(), output_collection)
+            if hasattr(tensor, "numpy"):
+                self._save_tensor_to_file(export_name, tensor.numpy(), output_collection)
 
     def _save_tensors_post_step(self, batch, logs):
         # some tensors available as value from within hook are saved here
@@ -1023,7 +1028,10 @@ def save_tape_logs(self, model_inputs=None, outputs=None):
         :return:
         """
         logs = {ModelOutput.PREDICTIONS: outputs, ModelInput.INPUTS: model_inputs}
-        self.save_smdebug_logs(logs)
+        if is_tf_version_2x() and tf.executing_eagerly():
+            self.save_smdebug_logs(logs)
+        else:
+            self.logger.warn("cannot save model inputs and outputs in non-eager execution mode")
 
     def wrap_tape(self, tape):
         """
diff --git a/tests/tensorflow2/test_grad_tape_tf_function.py b/tests/tensorflow2/test_grad_tape_tf_function.py
new file mode 100644
index 000000000..681287b14
--- /dev/null
+++ b/tests/tensorflow2/test_grad_tape_tf_function.py
@@ -0,0 +1,81 @@
+# Third Party
+import tensorflow as tf
+
+# First Party
+import smdebug.tensorflow as smd
+from smdebug.core.collection import CollectionKeys
+
+
+def create_hook(trial_dir):
+    hook = smd.KerasHook(trial_dir, save_all=True)
+    return hook
+
+
+def create_model():
+    model = tf.keras.models.Sequential(
+        [
+            # WA for TF issue https://github.com/tensorflow/tensorflow/issues/36279
+            tf.keras.layers.Flatten(input_shape=(28, 28, 1)),
+            tf.keras.layers.Dense(128, activation="relu"),
+            tf.keras.layers.Dropout(0.2),
+            tf.keras.layers.Dense(10, activation="softmax"),
+        ]
+    )
+    return model
+
+
+def test_gradtape_tf_function(out_dir):
+    def get_grads(images, labels):
+        # with tf.GradientTape() as tape:
+        return model(images, training=True)
+
+    @tf.function
+    def train_step(images, labels):
+        return tf.reduce_mean(get_grads(images, labels))
+
+    mnist = tf.keras.datasets.mnist
+    (x_train, y_train), _ = mnist.load_data()
+    dataset = tf.data.Dataset.from_tensor_slices(
+        (tf.cast(x_train[..., tf.newaxis] / 255, tf.float32), tf.cast(y_train, tf.int64))
+    )
+    dataset = dataset.shuffle(1000).batch(64)
+    model = create_model()
+    hook = create_hook(out_dir)
+    opt = tf.keras.optimizers.Adam()
+    hook.wrap_optimizer(opt)
+
+    n_epochs = 1
+    for epoch in range(n_epochs):
+        for data, labels in dataset:
+            dataset_labels = labels
+            labels = tf.one_hot(labels, depth=10)
+            with hook.wrap_tape(tf.GradientTape()) as tape:
+                logits = train_step(data, labels)
+            grads = tape.gradient(logits, model.variables)
+            opt.apply_gradients(zip(grads, model.variables))
+            hook.save_tensor("inputs", data, CollectionKeys.INPUTS)
+            hook.save_tensor("logits", logits, CollectionKeys.OUTPUTS)
+            hook.save_tensor("labels", labels, CollectionKeys.OUTPUTS)
+
+    model.save(out_dir, save_format="tf")
+    hook.close()
+
+    trial = smd.create_trial(out_dir)
+    assert trial.tensor_names(collection=CollectionKeys.LOSSES) == ["loss"]
+    assert trial.tensor_names(collection=CollectionKeys.WEIGHTS) == [
+        "weights/dense/kernel:0",
+        "weights/dense_1/kernel:0",
+    ]
+    assert trial.tensor_names(collection=CollectionKeys.BIASES) == [
+        "weights/dense/bias:0",
+        "weights/dense_1/bias:0",
+    ]
+    assert trial.tensor_names(collection=CollectionKeys.OPTIMIZER_VARIABLES) == [
+        "Adam/beta_1:0",
+        "Adam/beta_2:0",
+        "Adam/decay:0",
+        "Adam/iter:0",
+        "Adam/learning_rate:0",
+    ]
+    assert trial.tensor_names(collection=CollectionKeys.INPUTS) == ["inputs"]
+    assert trial.tensor_names(collection=CollectionKeys.OUTPUTS) == ["labels", "logits"]