How to properly use tensor_parallel while applying also Zero Stage 3

We were trying to pass the deepspeed config file to another training framework. We found that the configuration of tensor parallelism with Zero Stage 3 enabled is a bit confusing. For instance, if we set:

···
"tensor_parallel": {
    "autotp_size": 2
}

we would get the error

AssertionError Currently, the compatibility between autotp and zero_stage = 3 has been validated

To enable both TP and Zero 3, we would have to set:

···
"tensor_parallel": {
    "tensor_parallel": {"tp_size: 2}
}

We found the following commend in the code:

class TPConfig(DeepSpeedConfigModel):
""" Configure tensor parallelism settings """

tp_size: int = 1
""" Number of devices to split the model across using tensor parallelism. """

tp_grain_size: int = 1
"The variable required by the autoTP parser has not been activated in training yet"
"as it depends on the gather logic that supports uneven partitioning. "
"Desired MLP/lm_head tp size granularity. DNN library favors tensor size in granularity of power of 2, we pick  64 as a default size."

mpu: object = None
"""
A model parallelism unit object that implements
``get_{model,data}_parallel_{rank,group,world_size}()``.
"""

tp_group: object = None

What is the difference between this autotp_size and tp_size? Why autotp_size is not compatible with Zero 3? How should we properly avoid reduce GPU memory while Zero 3 still triggers OOM?