Allow freezing of FunctionGraph for hashing

pytensor/graph/basic.py

@@ -2,6 +2,7 @@
 
 import abc
 import warnings
+import weakref
 from collections.abc import (
     Hashable,
     Iterable,
@@ -14,6 +15,7 @@
     Any,
     Generic,
     Optional,
+    Self,
     TypeVar,
     Union,
     cast,
@@ -838,6 +840,93 @@ def value(self):
         return self.data
 
 
+def _get_frozen_output(apply_node: "FrozenApply", index: int) -> Variable:
+    """Resolve a FrozenApply output by index. Used by pickle."""
+    return apply_node.outputs[index]
+
+
+def _make_frozen_output_reduce(out: Variable):
+    """Create a __reduce_ex__ override for a FrozenApply output Variable."""
+    owner = out.owner
+    index = out.index
+
+    def __reduce_ex__(protocol):
+        return (_get_frozen_output, (owner, index))
+
+    return __reduce_ex__
+
+
+class FrozenApply(Apply):
+    """An immutable, globally-interned Apply node for frozen graphs.
+
+    Uses tuples for ``inputs`` and ``outputs`` so mutation raises ``TypeError``
+    at the language level.  Interned by ``(op, cache_key(inputs))`` —
+    constructing a ``FrozenApply`` with the same op and input variables returns
+    the cached instance.
+
+    Constants are keyed by ``(type, data_bytes)`` so that two independently
+    created Constants with the same value resolve to the same cached node.
+    """
+
+    _cache: weakref.WeakValueDictionary = weakref.WeakValueDictionary()
+
+    @staticmethod
+    def _input_to_key(inp: Variable):
+        """Convert an input Variable to a hashable, value-based cache key element.
+
+        Non-Constants (NominalVariables, FrozenApply outputs) are already
+        globally interned, so identity works.  Constants use their byte
+        representation so that independently-created equal constants
+        (including NaN) produce the same key.  Object-dtype constants
+        (e.g. slices) fall back to ``signature()`` since their byte
+        representation stores pointers, not values.
+        """
+        if isinstance(inp, Constant):
+            a = np.asarray(inp.data)
+            if a.dtype.kind != "O":
+                return (inp.type, a.tobytes(), a.dtype.str, a.shape)
+            return inp.signature()
+        return inp
+
+    def __new__(
+        cls,
+        op: "Op",
+        inputs: tuple[Variable, ...],
+        output_types: tuple["Type", ...],
+    ):
+        cache_key = (op, tuple(cls._input_to_key(i) for i in inputs))
+        cached = cls._cache.get(cache_key)
+        if cached is not None:
+            return cached
+
+        instance = object.__new__(cls)
+        instance.op = op
+        instance.inputs = inputs  # type: ignore[assignment]
+        instance.outputs = tuple(  # type: ignore[assignment]
+            t.variable_type(type=t, owner=instance, index=i)
+            for i, t in enumerate(output_types)
+        )
+        # Give each output Variable a __reduce__ that resolves to the
+        # canonical output on unpickle, avoiding fresh Variable objects.
+        for out in instance.outputs:
+            out.__reduce_ex__ = _make_frozen_output_reduce(out)  # type: ignore[method-assign]
+        instance.tag = Scratchpad()
+        cls._cache[cache_key] = instance
+        return instance
+
+    def __init__(self, op, inputs, output_types):
+        # All initialization is done in __new__
+        pass
+
+    def clone(self, clone_inner_graph: bool = False) -> Self:
+        """Frozen nodes are immutable — cloning returns self."""
+        return self
+
+    def __reduce__(self):
+        output_types = tuple(o.type for o in self.outputs)
+        return (type(self), (self.op, self.inputs, output_types))
+
+
 def clone(
     inputs: Sequence[Variable],
     outputs: Sequence[Variable],
@@ -1154,14 +1243,14 @@ def equal_computations(
 
     for x, y in zip(xs, ys, strict=True):
         if not isinstance(x, Variable) and not isinstance(y, Variable):
-            return np.array_equal(x, y)
+            return np.array_equal(x, y, equal_nan=True)
         if not isinstance(x, Variable):
             if isinstance(y, Constant):
-                return np.array_equal(y.data, x)
+                return np.array_equal(y.data, x, equal_nan=True)
             return False
         if not isinstance(y, Variable):
             if isinstance(x, Constant):
-                return np.array_equal(x.data, y)
+                return np.array_equal(x.data, y, equal_nan=True)
             return False
         x_is_owned, y_is_owned = (x.owner is not None, y.owner is not None)
         if x_is_owned != y_is_owned:

pytensor/graph/fg.py

@@ -1,6 +1,7 @@
 """A container for specifying and manipulating a graph with distinct inputs and outputs."""
 
 import time
+from abc import ABC, abstractmethod
 from collections import defaultdict
 from collections.abc import Iterable, Sequence
 from typing import Any, Union, cast
@@ -10,6 +11,8 @@
 from pytensor.graph.basic import (
     Apply,
     AtomicVariable,
+    Constant,
+    NominalVariable,
     Variable,
     clone_get_equiv,
 )
@@ -23,12 +26,25 @@
     toposort_with_orderings,
     vars_between,
 )
-from pytensor.graph.utils import MetaObject, MissingInputError, TestValueError
+from pytensor.graph.utils import MissingInputError, TestValueError
 
 
 ClientType = tuple[Apply, int]
 
 
+class AbstractFunctionGraph(ABC):
+    """Read-only interface shared by FunctionGraph and FrozenFunctionGraph."""
+
+    inputs: Sequence[Variable]
+    outputs: Sequence[Variable]
+    apply_nodes: set[Apply]
+    variables: set[Variable]
+    clients: dict[Variable, list[ClientType]]
+
+    @abstractmethod
+    def toposort(self) -> list[Apply]: ...
+
+
 class Output(Op):
     """A dummy `Op` that represents an output variable in a `FunctionGraph`."""
 
@@ -47,7 +63,7 @@ def __str__(self):
         return f"output[{self.idx}]"
 
 
-class FunctionGraph(MetaObject):
+class FunctionGraph(AbstractFunctionGraph):
     r"""
     A `FunctionGraph` represents a subgraph bound by a set of input variables and
     a set of output variables, ie a subgraph that specifies an PyTensor function.
@@ -928,3 +944,167 @@ def dprint(self, **kwargs):
         from pytensor.printing import debugprint
 
         return debugprint(self, **kwargs)
+
+    def freeze(self) -> "FrozenFunctionGraph":
+        """Return a frozen, hashable version of this FunctionGraph."""
+        return FrozenFunctionGraph(self.inputs, self.outputs)
+
+
+class FrozenFunctionGraph(AbstractFunctionGraph):
+    """Immutable, hashable function graph for inner graphs of Ops.
+
+    All internal nodes are globally interned via ``FrozenApply``.  Two
+    ``FrozenFunctionGraph`` instances built from structurally identical source
+    graphs share the same interned output objects, so equality reduces to
+    identity comparison on the outputs tuple.
+
+    Use ``FunctionGraph.freeze()`` or ``FrozenFunctionGraph(inputs, outputs)``
+    to create instances.
+
+    .. code-block:: python
+
+        from pytensor.scalar.basic import float64, add
+        from pytensor.graph.fg import FunctionGraph
+
+        x, y = float64("x"), float64("y")
+        frozen = FunctionGraph([x, y], [add(x, y)]).freeze()
+        frozen2 = FunctionGraph([x, y], [add(x, y)]).freeze()
+
+        assert frozen == frozen2
+        assert {frozen: "value"}[frozen2] == "value"
+    """
+
+    def __init__(
+        self,
+        inputs: Sequence[Variable],
+        outputs: Sequence[Variable],
+    ):
+        from pytensor.graph.basic import FrozenApply
+
+        nominal_inputs = tuple(
+            NominalVariable(i, inp.type, name=inp.name) for i, inp in enumerate(inputs)
+        )
+
+        memo: dict[Variable, Variable] = dict(zip(inputs, nominal_inputs, strict=True))
+
+        for node in toposort(outputs, blockers=inputs):
+            for inp in node.inputs:
+                if inp not in memo:
+                    if isinstance(inp, Constant):
+                        memo[inp] = inp
+                    elif isinstance(inp, AtomicVariable):
+                        memo[inp] = inp
+                    else:
+                        raise ValueError(
+                            f"Non-Constant, non-AtomicVariable orphan {inp} found "
+                            "in the graph. All variables must be graph inputs, "
+                            "Constants, AtomicVariables, or produced by Apply "
+                            "nodes reachable from the inputs."
+                        )
+
+            new_inputs = tuple(memo[i] for i in node.inputs)
+            output_types = tuple(out.type for out in node.outputs)
+            new_node = FrozenApply(node.op, new_inputs, output_types)
+
+            memo.update(zip(node.outputs, new_node.outputs, strict=True))
+
+        # Handle outputs that are Constants or AtomicVariables not
+        # encountered during toposort (e.g. a graph with no Apply nodes)
+        for o in outputs:
+            if o not in memo:
+                if isinstance(o, Constant):
+                    memo[o] = o
+                elif isinstance(o, AtomicVariable):
+                    memo[o] = o
+
+        try:
+            frozen_outputs = tuple(memo[o] for o in outputs)
+        except KeyError:
+            unmapped = [o for o in outputs if o not in memo]
+            raise ValueError(
+                f"Output variable {unmapped[0]} could not be mapped to a frozen "
+                "graph variable. All outputs must be graph inputs, "
+                "constants, or produced by Apply nodes reachable from "
+                "the inputs."
+            )
+
+        self.inputs: tuple[Variable, ...] = nominal_inputs
+        self.outputs: tuple[Variable, ...] = frozen_outputs
+        self._variables: set[Variable] | None = None
+        self._apply_nodes: set[Apply] | None = None
+        self._clients: dict[Variable, list[ClientType]] | None = None
+        self._toposort: list[Apply] | None = None
+
+    def __reduce__(self):
+        return FrozenFunctionGraph, (self.inputs, self.outputs)
+
+    def __hash__(self):
+        return hash(self.outputs)
+
+    def __eq__(self, other):
+        if self is other:
+            return True
+        if not isinstance(other, FrozenFunctionGraph):
+            return False
+        return self.inputs == other.inputs and self.outputs == other.outputs
+
+    def __repr__(self):
+        return f"FrozenFunctionGraph(inputs={list(self.inputs)}, outputs={list(self.outputs)})"
+
+    def __copy__(self):
+        return self
+
+    def __deepcopy__(self, memo):
+        return self
+
+    @property
+    def apply_nodes(self) -> set[Apply]:  # type: ignore[override]
+        if self._apply_nodes is None:
+            self._apply_nodes = set(applys_between(self.inputs, self.outputs))
+        return self._apply_nodes
+
+    def toposort(self) -> list[Apply]:
+        if self._toposort is None:
+            self._toposort = list(toposort(self.outputs, blockers=self.inputs))
+        return self._toposort
+
+    @property
+    def variables(self) -> set[Variable]:  # type: ignore[override]
+        if self._variables is None:
+            self._variables = set(vars_between(self.inputs, self.outputs))
+        return self._variables
+
+    @property
+    def clients(self) -> dict[Variable, list[ClientType]]:  # type: ignore[override]
+        if self._clients is None:
+            clients: dict[Variable, list[ClientType]] = {v: [] for v in self.inputs}
+            for node in self.toposort():
+                for i, inp in enumerate(node.inputs):
+                    clients.setdefault(inp, []).append((node, i))
+                for out in node.outputs:
+                    clients.setdefault(out, [])
+            self._clients = clients
+        return self._clients
+
+    def unfreeze(self) -> "FunctionGraph":
+        """Return a mutable FunctionGraph with fresh mutable Apply nodes."""
+        memo: dict[Variable, Variable] = {inp: inp.type() for inp in self.inputs}
+
+        for node in self.toposort():
+            for i in node.inputs:
+                if i not in memo:
+                    if isinstance(i, AtomicVariable):
+                        memo[i] = i
+                    else:
+                        memo[i] = i.clone()
+            new_inputs = [memo[i] for i in node.inputs]
+            new_node = Apply(
+                node.op,
+                new_inputs,
+                [o.type() for o in node.outputs],
+            )
+            memo.update(zip(node.outputs, new_node.outputs))
+
+        new_inputs = [memo[i] for i in self.inputs]
+        new_outputs = [memo[o] for o in self.outputs]
+        return FunctionGraph(new_inputs, new_outputs, clone=False)

pytensor/graph/op.py

@@ -25,7 +25,7 @@
 
 if TYPE_CHECKING:
     from pytensor.compile.function.types import Function
-    from pytensor.graph.fg import FunctionGraph
+    from pytensor.graph.fg import AbstractFunctionGraph, FunctionGraph
     from pytensor.graph.type import Type
 
 StorageCellType = list[Any | None]
@@ -459,16 +459,18 @@ def perform(
         """
 
     def do_constant_folding(self, fgraph: "FunctionGraph", node: Apply) -> bool:
-        """Determine whether or not constant folding should be performed for the given node.
+        """Determine whether constant folding should be performed for the given node.
 
         This allows each `Op` to determine if it wants to be constant
         folded when all its inputs are constant. This allows it to choose where
         it puts its memory/speed trade-off. Also, it could make things faster
-        as constants can't be used for in-place operations (see
-        ``*IncSubtensor``).
+        as constants can't be used for in-place operations (see ``*IncSubtensor``).
 
         Parameters
         ----------
+        fgraph : FunctionGraph
+            Function graph to which `node` belongs. This is passed in case the `Op` needs to inspect the graph to make
+            its decision.
         node : Apply
             The node for which the constant folding determination is made.
 
@@ -633,8 +635,8 @@ def perform(self, node, inputs, output_storage):
 class HasInnerGraph(ABC):
     r"""A mixin for an `Op` that contain an inner graph."""
 
-    fgraph: "FunctionGraph"
-    """A `FunctionGraph` of the inner function."""
+    fgraph: "AbstractFunctionGraph"
+    """The inner function graph (FunctionGraph or FrozenFunctionGraph)."""
 
     @property
     @abstractmethod