Access to model.logp changes the free random variables inplace

[aseyboldt]

with pm.Model() as model:
    a = pm.HalfNormal("a")
    c = pm.Uniform("c", lower=a - 1, upper=1)

print("before")
print("------")
aesara.dprint(c)

print()
model.logpt

print("after")
print("-----")
aesara.dprint(c)

before
------
uniform_rv{0, (0, 0), floatX, False}.1 [id A] 'c'   
 |RandomStateSharedVariable(<RandomState(MT19937) at 0x7F649E351540>) [id B]
 |TensorConstant{[]} [id C]
 |TensorConstant{11} [id D]
 |Elemwise{sub,no_inplace} [id E] ''   
 | |halfnormal_rv{0, (0, 0), floatX, False}.1 [id F] 'a'   
 | | |RandomStateSharedVariable(<RandomState(MT19937) at 0x7F649D753040>) [id G]
 | | |TensorConstant{[]} [id H]
 | | |TensorConstant{11} [id I]
 | | |TensorConstant{0.0} [id J]
 | | |TensorConstant{1.0} [id K]
 | |TensorConstant{1} [id L]
 |TensorConstant{1.0} [id M]

after
-----
uniform_rv{0, (0, 0), floatX, False}.1 [id A] 'c'   
 |RandomStateSharedVariable(<RandomState(MT19937) at 0x7F649E351540>) [id B]
 |TensorConstant{[]} [id C]
 |TensorConstant{11} [id D]
 |Elemwise{sub,no_inplace} [id E] ''   
 | |Elemwise{exp,no_inplace} [id F] ''   
 | | |a_log__ [id G]
 | |TensorConstant{1} [id H]
 |TensorConstant{1.0} [id I]

Maybe some graph.replace_all is changing the variables inplace when it should have made a copy first?

CC @brandonwillard

[aseyboldt]

This does not seem to be an issue in the aeppl implementation, so maybe we can just close this when #4887 is merged.

This does not seem to change the original variables:

import aeppl

with pm.Model() as model:
    a = pm.Normal("a", sigma=1)
    b = pm.Normal("y", mu=a, sigma=1, observed=2)
    c = pm.Normal("z", mu=a + b)

aesara.dprint(b)
print()
aeppl.joint_logprob(model.rvs_to_values)
print()
aesara.dprint(b)