Fix bug with chained CustomSymbolicDists

Author: ricardoV94

pymc/distributions/distribution.py

@@ -27,7 +27,7 @@
 
 from pytensor import tensor as pt
 from pytensor.compile.builders import OpFromGraph
-from pytensor.graph import FunctionGraph, clone_replace, node_rewriter
+from pytensor.graph import FunctionGraph, graph_replace, node_rewriter
 from pytensor.graph.basic import Apply, Variable
 from pytensor.graph.rewriting.basic import in2out
 from pytensor.graph.utils import MetaType
@@ -588,7 +588,9 @@ def inline_symbolic_random_variable(fgraph, node):
     """Expand a SymbolicRV when obtaining the logp graph if `inline_logprob` is True."""
     op = node.op
     if op.inline_logprob:
-        return clone_replace(op.inner_outputs, dict(zip(op.inner_inputs, node.inputs)))
+        return graph_replace(
+            op.inner_outputs, dict(zip(op.inner_inputs, node.inputs)), strict=False
+        )
 
 
 # Registered before pre-canonicalization which happens at position=-10

tests/distributions/test_custom.py

@@ -21,6 +21,7 @@
 from numpy import random as npr
 from pytensor import scan
 from pytensor import tensor as pt
+from pytensor.graph import FunctionGraph
 from scipy import stats as st
 
 from pymc.distributions import (
@@ -42,11 +43,11 @@
     Uniform,
 )
 from pymc.distributions.custom import CustomDist, CustomDistRV, CustomSymbolicDistRV
-from pymc.distributions.distribution import support_point
+from pymc.distributions.distribution import inline_symbolic_random_variable, support_point
 from pymc.distributions.shape_utils import change_dist_size, rv_size_is_none, to_tuple
 from pymc.distributions.transforms import log
 from pymc.exceptions import BlockModelAccessError
-from pymc.logprob import logcdf, logp
+from pymc.logprob import conditional_logp, logcdf, logp
 from pymc.model import Deterministic, Model
 from pymc.pytensorf import collect_default_updates
 from pymc.sampling import draw, sample, sample_posterior_predictive
@@ -648,3 +649,71 @@ def dist(p, size):
         assert out.owner.op.extended_signature == "[size],(),[rng]->(),[rng]"
         assert out.owner.op.ndim_supp == 0
         assert out.owner.op.ndims_params == [0]
+
+    def test_inline_does_not_duplicate_graph(self):
+        mu = Normal.dist()
+        x = CustomDist.dist(mu, dist=lambda mu, size: Normal.dist(mu, size=size))
+
+        fgraph = FunctionGraph(outputs=[x], clone=False)
+        [inner_x, inner_rng_update] = inline_symbolic_random_variable.transform(fgraph, x.owner)
+        assert inner_rng_update.owner.inputs[-2] is mu
+        assert inner_x.owner.inputs[-2] is mu
+
+    def test_chained_custom_dist_bug(self):
+        """Regression test for issue reported in https://discourse.pymc.io/t/error-with-custom-distribution-after-using-scan/16255
+
+        This bug was caused by a duplication of a Scan-based CustomSymbolicDist when inlining another CustomSymbolicDist that used it as an input.
+        PyTensor failed to merge the two Scan graphs, causing a failure in the logp extraction.
+        """
+
+        rng = np.random.default_rng(123)
+        steps = 4
+        batch = 2
+
+        def scan_dist(seq, n_steps, size):
+            def step(s):
+                innov = Normal.dist()
+                traffic = s + innov
+                return traffic, {innov.owner.inputs[0]: innov.owner.outputs[0]}
+
+            rv_seq, _ = pytensor.scan(
+                fn=step,
+                sequences=[seq],
+                outputs_info=[None],
+                n_steps=n_steps,
+                strict=True,
+            )
+            return rv_seq
+
+        def normal_shifted(mu, size):
+            return Normal.dist(mu=mu, size=size) - 1
+
+        seq = pt.matrix("seq", shape=(batch, steps))
+        latent_rv = CustomDist.dist(
+            seq.T,
+            steps,
+            dist=scan_dist,
+            shape=(steps, batch),
+        )
+        latent_rv.name = "latent"
+
+        observed_rv = CustomDist.dist(
+            latent_rv,
+            dist=normal_shifted,
+            shape=(steps, batch),
+        )
+        observed_rv.name = "observed"
+
+        latent_vv = latent_rv.type()
+        observed_vv = observed_rv.type()
+
+        observed_logp = conditional_logp({latent_rv: latent_vv, observed_rv: observed_vv})[
+            observed_vv
+        ]
+        latent_vv_test = rng.standard_normal(size=(steps, batch))
+        observed_vv_test = rng.standard_normal(size=(steps, batch))
+        expected_logp = st.norm.logpdf(observed_vv_test + 1, loc=latent_vv_test)
+        np.testing.assert_allclose(
+            observed_logp.eval({latent_vv: latent_vv_test, observed_vv: observed_vv_test}),
+            expected_logp,
+        )