Tensor flow (RT) RAM leakeage during inference

Question

I have multipole calls to:

predictions = self.inference_function(**{self.input_tensor_name: tf.constant(input_image, dtype=tf.float32)})[self.output_tensor_name].numpy()

Each call my RAM (not in GPU) is increased in small amount. After some times this throw me out of my script.

This is how I load my saved model:

    trt_saved_model = tf.saved_model.load(model_path)
    inference_function = trt_saved_model.signatures["serving_default"]
    input_tensor_name = list(inference_function.structured_input_signature[1].keys())[0]
    output_tensor_name = list(inference_function.structured_outputs.keys())[0]

Am I doing something wrong?