tfp.distributions.RelaxedBernoulli produces samples on the boundary with inf log_prob

When using relaxed_bernouli to sample on boundary, it outputs inf or NaN log_prob.

import tensorflow as tf
import tensorflow_probability as tfp
tf.random.set_seed(2)
dist = tfp.distributions.RelaxedBernoulli(temperature = tf.constant(0.005), logits = tf.constant(-5.0))
sample = dist.sample()
print('Sample: {}'.format(sample))  # should be non-zero value
print('Log prob: {}'.format(dist.log_prob(sample).numpy()))  # should be non-inf value


dist = tfp.distributions.RelaxedBernoulli(temperature = tf.constant(0.0), logits = tf.constant(-5.0))
sample = dist.sample()
print('Sample: {}'.format(sample))  # should be non-nan value
print('Log prob: {}'.format(dist.log_prob(sample).numpy()))  # should be non-nan value

Outputs:

Sample: 0.0
Log prob: inf
Sample: nan
Log prob: nan

Expected Outputs:

Sample: normal value but not not 0.0
Log prob: normal value not not inf
Sample: something not nan
Log prob: something not nan

Start by running the supplied RelaxedBernoulli reproduction with temperatures 0.005 and 0.0, then inspect the RelaxedBernoulli sampling and log_prob implementation. Determine how boundary or non-finite samples arise and verify that both examples produce finite, non-NaN samples and log probabilities.