Poisson random variable samples float instead of int, fails in hierarchical modeling as result

System information

• Have I written custom code (as opposed to using a stock example script provided in TensorFlow): Yes, derived from tf example.
• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): MacOS Mojave 10.14.4
• TensorFlow installed from (source or binary): pip
• TensorFlow version (use command below): v1.14.0-rc1-22-gaf24dc91b5 1.14.0 and 2.0
• Python version: Python 3.7.3
• Exact command to reproduce: I am implementing a hierarchical model, where the value of the Poisson sample is used to determine the batch size of the normal distribution. Since the Poisson sample is a float, the MCMC step function treats it as a float. This results in incorrect mathematics for MCMC to work.

import tensorflow as tf
import tensorflow_probability as tfp
tfd = tfp.distributions
tfb = tfp.bijectors

rv_m = tfd.Poisson(rate=1.)
m    = rv_m.sample()

def joint_log_prob(m):
    rv_m    = tfd.Poisson(rate=1.)
    m_int   = tf.cast(m,'int32')
    rv_norm = tfd.Normal(loc=tf.zeros(m_int),scale=1.)
    return (tf.constant(-1.))

number_of_steps = 1000
burnin          = 0

initial_chain_state      = [m]
log_prob                 = lambda *args: joint_log_prob(*args)
[post_m], kernel_results = tfp.mcmc.sample_chain(num_results=number_of_steps,num_burnin_steps=burnin,current_state=initial_chain_state,kernel=tfp.mcmc.RandomWalkMetropolis(target_log_prob_fn=log_prob,seed=4))

Reproduce the supplied Python snippet, starting with tfd.Poisson.sample and the tfp.mcmc.RandomWalkMetropolis call. Trace where the Poisson sample's dtype enters the hierarchical model and MCMC state. Done means the reproduced model preserves the intended integer-valued Poisson state and its MCMC step uses it correctly.