Proper way to add custom assets to Keras model? - tensorflow

Question

I want to include arbitrary assets to my saved_model (for example, a text file/json mapping label indices to their proper names). Historically, I just manually save the assets to {saved_model_path}/assets/custom_asset.txt.
Today I came across tf.saved_model.Asset which seems like a clean way to do this:
checkpoint = tf.train.Checkpoint()
checkpoint.my_filename = tf.saved_model.Asset('some_file.txt')
# Normal model saving stuff
model.save(saved_model_path)
# Save tf.saved_model.Asset
tf.saved_model.save(checkpoint, saved_model_path)
The above seems to save the model fine, plus the custom asset shows up in the {saved_model}/assets directory as expected. Clearly I'm doing something wrong:
Using saved_model_cli to inspect the model shows different signatures and such compared to doing the same thing, but not saving the custom trackable object.
Loading the model fails:
# Attempting to load the model in a separate process
model = tf.keras.models.load_model(path)
>>> Traceback (most recent call last):
File "/Users/rsilveira/dev/ddm-ml-classifieds-recommender/tf_model_exporter.py", line 80, in <module>
load(_saved_model_path)
File "/Users/rsilveira/dev/ddm-ml-classifieds-recommender/tf_model_exporter.py", line 67, in load
model = tf.keras.models.load_model(path)
File "/Users/rsilveira/dev/ddm-ml-classifieds-recommender/.venv/lib/python3.8/site-packages/keras/utils/traceback_utils.py", line 67, in error_handler
raise e.with_traceback(filtered_tb) from None
File "/Users/rsilveira/dev/ddm-ml-classifieds-recommender/.venv/lib/python3.8/site-packages/keras/saving/saved_model/load.py", line 466, in _revive_from_config
obj, self._proto.nodes[node_id], node_id)
IndexError: list index (62) out of range
Question
Is there a "native" Tensorflow-y way to save custom assets (not attached to the model)?
Reproducible code
# Create some dummy model
def build_model():
i = Input(shape=(5,))
fc = Dense(5, activation='relu')(i)
softmax = Dense(5, activation='softmax', name='Softmax')(fc)
return Model(inputs=i, outputs=softmax)
# Some dummy model
model = build_model()
opt = tf.keras.optimizers.Adam(learning_rate=0.003)
model.compile(optimizer=opt,
metrics=[SparseCategoricalAccuracy()],
loss=SparseCategoricalCrossentropy(from_logits=False))
# Dummy data
_ds = Dataset.from_tensor_slices((np.random.rand(10, 5), np.random.randint(0, 5, 10)))
model.fit(_ds.batch(32), epochs=3)
# Save the model
model.save(saved_model_path)
some_file = 'some_file.txt'
checkpoint = tf.train.Checkpoint()
checkpoint.my_filename = tf.saved_model.Asset(some_file)
tf.saved_model.save(checkpoint, saved_model_path)
# Loading the model - in a different process
model = tf.keras.models.load_model(saved_model_path)