Fix copying of shared variables in fgraph_from_model (#7153)

* Do not use deprecated ScalarSharedVariable

* Recreate SharedVariables with exact type in fgraph_from_model

pymc/model/core.py

@@ -44,7 +44,6 @@
 from pytensor.tensor.elemwise import Elemwise
 from pytensor.tensor.random.op import RandomVariable
 from pytensor.tensor.random.type import RandomType
-from pytensor.tensor.sharedvar import ScalarSharedVariable
 from pytensor.tensor.variable import TensorConstant, TensorVariable
 from typing_extensions import Self
 
@@ -999,6 +998,7 @@ def add_coord(
                 length = pytensor.shared(length, name=name)
             else:
                 length = pytensor.tensor.constant(length)
+        assert length.type.ndim == 0
         self._dim_lengths[name] = length
         self._coords[name] = values
 
@@ -1028,7 +1028,7 @@ def set_dim(self, name: str, new_length: int, coord_values: Optional[Sequence] =
         coord_values
             Optional sequence of coordinate values.
         """
-        if not isinstance(self.dim_lengths[name], ScalarSharedVariable):
+        if not isinstance(self.dim_lengths[name], SharedVariable):
             raise ValueError(f"The dimension '{name}' is immutable.")
         if coord_values is None and self.coords.get(name, None) is not None:
             raise ValueError(
@@ -1188,7 +1188,7 @@ def set_data(
                             actual=new_length,
                             expected=old_length,
                         )
-                if isinstance(length_tensor, ScalarSharedVariable):
+                if isinstance(length_tensor, SharedVariable):
                     # The dimension is mutable, but was defined without being linked
                     # to a shared variable. This is allowed, but a little less robust.
                     self.set_dim(dname, new_length, coord_values=new_coords)

pymc/model/fgraph.py

@@ -11,18 +11,17 @@
 #   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 #   See the License for the specific language governing permissions and
 #   limitations under the License.
-from copy import copy
+from copy import copy, deepcopy
 from typing import Optional
 
 import pytensor
 
-from pytensor import Variable, shared
+from pytensor import Variable
 from pytensor.compile import SharedVariable
 from pytensor.graph import Apply, FunctionGraph, Op, node_rewriter
 from pytensor.graph.rewriting.basic import out2in
 from pytensor.scalar import Identity
 from pytensor.tensor.elemwise import Elemwise
-from pytensor.tensor.sharedvar import ScalarSharedVariable
 
 from pymc.logprob.transforms import Transform
 from pymc.model.core import Model
@@ -113,6 +112,21 @@ def local_remove_identity(fgraph, node):
 remove_identity_rewrite = out2in(local_remove_identity)
 
 
+def deepcopy_shared_variable(var: SharedVariable) -> SharedVariable:
+    # Shared variables don't have a deepcopy method (SharedVariable.clone reuses the old container and contents).
+    # We recreate Shared Variables manually after deepcopying their container.
+    new_var = type(var)(
+        type=var.type,
+        value=None,
+        strict=None,
+        container=deepcopy(var.container),
+        name=var.name,
+    )
+    assert new_var.type == var.type
+    new_var.tag = copy(var.tag)
+    return new_var
+
+
 def fgraph_from_model(
     model: Model, inlined_views=False
 ) -> tuple[FunctionGraph, dict[Variable, Variable]]:
@@ -192,18 +206,7 @@ def fgraph_from_model(
     shared_vars_to_copy += [v for v in model.dim_lengths.values() if isinstance(v, SharedVariable)]
     shared_vars_to_copy += [v for v in model.named_vars.values() if isinstance(v, SharedVariable)]
     for var in shared_vars_to_copy:
-        # FIXME: ScalarSharedVariables are converted to 0d numpy arrays internally,
-        #  so calling shared(shared(5).get_value()) returns a different type: TensorSharedVariables!
-        #  Furthermore, PyMC silently ignores mutable dim changes that are SharedTensorVariables...
-        #  https://github.com/pymc-devs/pytensor/issues/396
-        if isinstance(var, ScalarSharedVariable):
-            new_var = shared(var.get_value(borrow=False).item())
-        else:
-            new_var = shared(var.get_value(borrow=False))
-
-        assert new_var.type == var.type
-        new_var.name = var.name
-        new_var.tag = copy(var.tag)
+        new_var = deepcopy_shared_variable(var)
         # We can replace input variables by placing them in the memo
         memo[var] = new_var
 

tests/model/test_core.py

@@ -35,7 +35,7 @@
 from pytensor.raise_op import Assert
 from pytensor.tensor import TensorVariable
 from pytensor.tensor.random.op import RandomVariable
-from pytensor.tensor.sharedvar import ScalarSharedVariable
+from pytensor.tensor.sharedvar import TensorSharedVariable
 from pytensor.tensor.variable import TensorConstant
 
 import pymc as pm
@@ -823,7 +823,7 @@ def test_add_coord_mutable_kwarg():
         m.add_coord("fixed", values=[1], mutable=False)
         m.add_coord("mutable1", values=[1, 2], mutable=True)
         assert isinstance(m._dim_lengths["fixed"], TensorConstant)
-        assert isinstance(m._dim_lengths["mutable1"], ScalarSharedVariable)
+        assert isinstance(m._dim_lengths["mutable1"], TensorSharedVariable)
         pm.MutableData("mdata", np.ones((1, 2, 3)), dims=("fixed", "mutable1", "mutable2"))
         assert isinstance(m._dim_lengths["mutable2"], TensorVariable)
 

tests/model/test_fgraph.py

@@ -107,14 +107,16 @@ def test_data(inline_views):
     with pm.Model(coords_mutable={"test_dim": range(3)}) as m_old:
         x = pm.MutableData("x", [0.0, 1.0, 2.0], dims=("test_dim",))
         y = pm.MutableData("y", [10.0, 11.0, 12.0], dims=("test_dim",))
+        sigma = pm.MutableData("sigma", [1.0], shape=(1,))
         b0 = pm.ConstantData("b0", np.zeros((1,)))
         b1 = pm.DiracDelta("b1", 1.0)
         mu = pm.Deterministic("mu", b0 + b1 * x, dims=("test_dim",))
-        obs = pm.Normal("obs", mu, sigma=1e-5, observed=y, dims=("test_dim",))
+        obs = pm.Normal("obs", mu=mu, sigma=sigma, observed=y, dims=("test_dim",))
 
     m_fgraph, memo = fgraph_from_model(m_old, inlined_views=inline_views)
     assert isinstance(memo[x].owner.op, ModelNamed)
     assert isinstance(memo[y].owner.op, ModelNamed)
+    assert isinstance(memo[sigma].owner.op, ModelNamed)
     assert isinstance(memo[b0].owner.op, ModelNamed)
     mu_inp = memo[mu].owner.inputs[0]
     obs = memo[obs]
@@ -124,10 +126,13 @@ def test_data(inline_views):
         assert mu_inp.owner.inputs[1].owner.inputs[1] is memo[x].owner.inputs[0]
         # ObservedRV(obs, y, *dims) not ObservedRV(obs, Named(y), *dims)
         assert obs.owner.inputs[1] is memo[y].owner.inputs[0]
+        # ObservedRV(Normal(..., sigma), ...) not ObservedRV(Normal(..., Named(sigma)), ...)
+        assert obs.owner.inputs[0].owner.inputs[4] is memo[sigma].owner.inputs[0]
     else:
         assert mu_inp.owner.inputs[0] is memo[b0]
         assert mu_inp.owner.inputs[1].owner.inputs[1] is memo[x]
         assert obs.owner.inputs[1] is memo[y]
+        assert obs.owner.inputs[0].owner.inputs[4] is memo[sigma]
 
     m_new = model_from_fgraph(m_fgraph)
 
@@ -140,9 +145,17 @@ def test_data(inline_views):
     # Shared model variables, dim lengths, and rngs are copied and no longer point to the same memory
     assert not same_storage(m_new["x"], x)
     assert not same_storage(m_new["y"], y)
+    assert not same_storage(m_new["sigma"], sigma)
     assert not same_storage(m_new["b1"].owner.inputs[0], b1.owner.inputs[0])
     assert not same_storage(m_new.dim_lengths["test_dim"], m_old.dim_lengths["test_dim"])
 
+    # Check they have the same type
+    assert m_new["x"].type == x.type
+    assert m_new["y"].type == y.type
+    assert m_new["sigma"].type == sigma.type
+    assert m_new["b1"].owner.inputs[0].type == b1.owner.inputs[0].type
+    assert m_new.dim_lengths["test_dim"].type == m_old.dim_lengths["test_dim"].type
+
     # Updating model shared variables in new model, doesn't affect old one
     with m_new:
         pm.set_data({"x": [100.0, 200.0]}, coords={"test_dim": range(2)})
@@ -155,22 +168,31 @@ def test_data(inline_views):
 @config.change_flags(floatX="float64")  # Avoid downcasting Ops in the graph
 def test_shared_variable():
     """Test that user defined shared variables (other than RNGs) aren't copied."""
-    x = shared(np.array([1, 2, 3.0]), name="x")
-    y = shared(np.array([1, 2, 3.0]), name="y")
+    mu = shared(np.array([1, 2, 3.0]), shape=(None,), name="mu")
+    sigma = shared(np.array([1.0]), shape=(1,), name="sigma")
+    obs = shared(np.array([1, 2, 3.0]), shape=(3,), name="obs")
 
     with pm.Model() as m_old:
-        test = pm.Normal("test", mu=x, observed=y)
+        test = pm.Normal("test", mu=mu, sigma=sigma, observed=obs)
 
-    assert test.owner.inputs[3] is x
-    assert m_old.rvs_to_values[test] is y
+    assert test.owner.inputs[3] is mu
+    assert test.owner.inputs[4] is sigma
+    assert m_old.rvs_to_values[test] is obs
 
     m_new = clone_model(m_old)
     test_new = m_new["test"]
     # Shared Variables are cloned but still point to the same memory
-    assert test_new.owner.inputs[3] is not x
-    assert m_new.rvs_to_values[test_new] is not y
-    assert same_storage(test_new.owner.inputs[3], x)
-    assert same_storage(m_new.rvs_to_values[test_new], y)
+    mu_new, sigma_new = test_new.owner.inputs[3:5]
+    obs_new = m_new.rvs_to_values[test_new]
+    assert mu_new is not mu
+    assert sigma_new is not sigma
+    assert obs_new is not obs
+    assert mu_new.type == mu.type
+    assert sigma_new.type == sigma.type
+    assert obs_new.type == obs.type
+    assert same_storage(mu, mu_new)
+    assert same_storage(sigma, sigma_new)
+    assert same_storage(obs, obs_new)
 
 
 @pytest.mark.parametrize("inline_views", (False, True))