ValueError: Found zero restored functions for caller function

I'm facing the same issue. Unfortunately I can't find a solution to the ValueError issue that avoids having to futz with some gnarly parts of the tensorflow library directly. I would suggest that the best way to solve completely will be to narrow down a working version of the tensorflow and keras dependencies with trial and error, as presumably the tutorial worked at some point! I haven't done this yet myself, but from github issues it looks like we're looking at before November of 2020 for a working set of dependencies.

However, if you do just need the weights, the following solution works for me:

import tensorflow as tf
from tensorflow.keras.layers.experimental import preprocessing

from rnn.simple_gru import SimpleGru # This class is nearly identical to MyModel in the tutorial
from predictors.one_step import StepPredictor # This class is nearly identical to OneStep in the tutorial, but change method name generate_one_step to method name predict_step

embedding_dim = 256
rnn_units = 1024

checkpoint_dir =  "training_checkpoints"
checkpoint_path = checkpoint_dir + "/ckpt_20"
print(os.listdir(checkpoint_dir))

corpus_file = ...file saved

tf.config.run_functions_eagerly(True)

text = open(corpus_file, 'rb').read().decode(encoding='utf-8')
vocab = sorted(set(text))
print(f'{len(vocab)} unique characters')

ids_from_chars = preprocessing.StringLookup(
    vocabulary=list(vocab), mask_token=None)
chars_from_ids = tf.keras.layers.experimental.preprocessing.StringLookup(
    vocabulary=ids_from_chars.get_vocabulary(), invert=True, mask_token=None)
def text_from_ids(ids):
    return tf.strings.reduce_join(chars_from_ids(ids), axis=-1)

model = SimpleGru(
    vocab_size=len(ids_from_chars.get_vocabulary()),
    embedding_dim=embedding_dim,
    rnn_units=rnn_units
)
model.built = True

adam = tf.keras.optimizers.Adam(
    learning_rate=tf.Variable(0.001),
    beta_1=tf.Variable(0.9),
    beta_2=tf.Variable(0.999),
    epsilon=tf.Variable(1e-7),
)
adam.iterations
adam.decay = tf.Variable(0.0)
loss = tf.losses.SparseCategoricalCrossentropy(from_logits=True)
model.compile(optimizer=adam, loss=loss)

load_status = model.load_weights(filepath=checkpoint_path)

def split_input_target(sequence):
    input_text = sequence[:-1]
    target_text = sequence[1:]
    return input_text, target_text

all_ids = ids_from_chars(tf.strings.unicode_split(text, 'UTF-8'))
ids_dataset = tf.data.Dataset.from_tensor_slices(all_ids)
seq_length = 100
sequences = ids_dataset.batch(seq_length+1, drop_remainder=True)
dataset = sequences.map(split_input_target)
loss = model.evaluate(dataset)
print("Restored model, loss: {:5.2f}".format(loss))

# Not necessary for this to work, but useful for debugging
load_status.assert_consumed()

step_predictor = StepPredictor(model, chars_from_ids, ids_from_chars)

states = None
next_char = tf.constant(['ROMEO:'])
result = [next_char]

for n in range(100):
  next_char, states = step_predictor.predict_step(next_char, states=states)
  result.append(next_char)

print(tf.strings.join(result)[0].numpy().decode('utf-8'))

While this should work, it takes a long time to load on larger models.

Trying to get_weights from the model does not work for me (AttributeError: '_UserObject' object has no attribute 'get_weights') - this is why I'm loading them from the last checkpoint.

I would personally suggest taking what you've learned from the tutorial and moving on to other things; not everything in the tutorial can work well and much of the behaviour it depends on is poorly documented, outdated or both. I don't think tensorflow follows semver at all either, as there are many breaking changes in v2, which makes things harder.