Here in doc strings of log_prob function method - https://github.com/sbi-dev/sbi/blob/main/sbi/neural_nets/estimators/nflows_flow.py#L84C46-L84C81

Args:
            input: Inputs to evaluate the log probability on. Of shape
                `(sample_dim, batch_dim, *event_shape)`.
            condition: Conditions of shape `(sample_dim, batch_dim, *event_shape)`.

Condition variable is expected to be of format (sample_dim, batch_dim, event_shape)
Dimension 0 - Sample
Dimension 1 - Batch
Dimension 2 - Event

But in function definition in this line - https://github.com/sbi-dev/sbi/blob/main/sbi/neural_nets/estimators/nflows_flow.py#L94

...
condition_batch_dim = condition.shape[0]
...

0th dimension is considered as batch dimension, while it should be 1st dimension as per the doc strings.

The mismatch might be lead to usage of incorrect tensor shapes.