Do not include rvs in symbolic normalizing constant

pymc/pytensorf.py

@@ -979,7 +979,7 @@ def constant_fold(
     """
     fg = FunctionGraph(outputs=xs, features=[ShapeFeature()], copy_inputs=False, clone=True)
 
-    # The default rewrite_graph includes a constand_folding that is not always applied.
+    # The default rewrite_graph includes a constant_folding that is not always applied.
     # We use an unconditional constant_folding as the last pass to ensure a thorough constant folding.
     rewrite_graph(fg)
     topo_unconditional_constant_folding.apply(fg)

pymc/variational/minibatch_rv.py

@@ -40,6 +40,9 @@ def make_node(self, rv, *total_size):
         out = rv.type()
         return Apply(self, [rv, *total_size], [out])
 
+    def infer_shape(self, fgraph, node, shapes):
+        return [shapes[0]]
+
     def perform(self, node, inputs, output_storage):
         output_storage[0][0] = inputs[0]
 

pymc/variational/opvi.py

@@ -74,6 +74,7 @@
 from pymc.pytensorf import (
     SeedSequenceSeed,
     compile,
+    constant_fold,
     find_rng_nodes,
     reseed_rngs,
 )
@@ -1105,7 +1106,10 @@ def symbolic_normalizing_constant(self):
         t = self.to_flat_input(
             pt.max(
                 [
-                    get_scaling(v.owner.inputs[1:], v.shape)
+                    get_scaling(
+                        v.owner.inputs[1:],
+                        constant_fold([v.owner.inputs[0].shape], raise_not_constant=False),
+                    )
                     for v in self.group
                     if isinstance(v.owner.op, MinibatchRandomVariable)
                 ]
@@ -1272,7 +1276,10 @@ def symbolic_normalizing_constant(self):
         t = pt.max(
             self.collect("symbolic_normalizing_constant")
             + [
-                get_scaling(obs.owner.inputs[1:], obs.shape)
+                get_scaling(
+                    obs.owner.inputs[1:],
+                    constant_fold([obs.owner.inputs[0].shape], raise_not_constant=False),
+                )
                 for obs in self.model.observed_RVs
                 if isinstance(obs.owner.op, MinibatchRandomVariable)
             ]

tests/variational/test_opvi.py

@@ -20,6 +20,7 @@
 
 import pymc as pm
 
+from pymc.testing import assert_no_rvs
 from pymc.variational import opvi
 from pymc.variational.approximations import (
     Empirical,
@@ -278,3 +279,18 @@ def test_logq_globals(three_var_approx):
     es = symbolic_logq.eval()
     assert e.shape == ()
     assert es.shape == (2,)
+
+
+def test_symbolic_normalizing_constant_no_rvs():
+    # Test that RVs aren't included in the graph of symbolic_normalizing_constant
+    rng = np.random.default_rng()
+
+    with pm.Model() as m:
+        obs = pm.Data("obs", rng.normal(size=(1000,)))
+        obs_batch = pm.Minibatch(obs, batch_size=128)
+        x = pm.Normal("x")  # Need at least one Free_RV in the graph
+        y_hat = pm.Flat("y_hat", observed=obs_batch, total_size=1000)
+
+        step = pm.ADVI()
+
+    assert_no_rvs(step.approx.symbolic_normalizing_constant)