MetricsLogger error for DreamerV3

RLlib Configure Algorithm, Training, Evaluation, Scaling
1

I am running DreamerV3 and facing some issues apparently due to the MetricsLogger, the raised error is similar to TypeError: Scalar tensor has no len(), and seems to be thrown when executing the _reduced_values() function on a 2D EagerTensor, in particular on these 2 lines of ray/rllib/utils/metrics/stats.py (e.g. when reduce_meth is tf.reduce_mean):

reduce_meth = getattr(tf, "reduce_" + self._reduce_method)
reduced = reduce_meth(values)

The error can be reproduced through the following code (I am using Ray 2.43.0 and tensorflow 2.19.0):

# See https://discuss.ray.io/t/dreamer-v3-rllib-tensorflow-error/15127
import os
os.environ["TF_USE_LEGACY_KERAS"] = "1"

import ray
from ray.rllib.algorithms.dreamerv3.dreamerv3 import DreamerV3Config

if __name__ == "__main__":

    ray.init(local_mode=True)  # debugging

    config = (
        DreamerV3Config()
        .environment(env='CartPole-v1')
        .env_runners(
            num_env_runners=0,
            num_envs_per_env_runner=1
        )
        .reporting(
            metrics_num_episodes_for_smoothing=1,
            report_images_and_videos=False,
            report_dream_data=False,
            report_individual_batch_item_stats=False,
        )
        # See Appendix A.
        .training(model_size="XS")
        .framework(
            "tf2",
            eager_tracing=False  # debugging
        )
    )

    agent = config.build_algo()
    agent.train()

Naively replacing reduce_meth(values) with reduce_meth(values[0]) seems to run but I do not think this is the ultimate solution. Is the above code the correct way to run DreamerV3?

Thank you