Fix inner graph inplace rewrites in Numba / PyTorch backends

pytensor/compile/function/types.py

@@ -5,6 +5,7 @@
 import logging
 import time
 import warnings
+from collections.abc import Sequence
 from itertools import chain
 from typing import TYPE_CHECKING
 
@@ -168,6 +169,59 @@
                 raise InconsistencyError(f"Trying to destroy a protected variable: {r}")
 
 
+def add_supervisor_to_fgraph(
+    fgraph: FunctionGraph,
+    input_specs: Sequence[SymbolicInput],
+    accept_inplace: bool = False,
+) -> None:
+    """Setup Supervisor Feature in a FunctionGraph, so that inplace rewrites can be used.
+
+    Parameters
+    ----------
+    fgraph: FunctionGraph
+        The FunctionGraph to setup the Supervisor Feature in.
+    input_specs: Sequence of SymbolicInput
+        The input specifications for the FunctionGraph.
+        Inputs with the attribute `mutable=False` and which are not already destroyed by an inplace operation
+        (if `accept_inplace` is True) will be protected from inplace operations.
+        Otherwise, they will be allowed to be destroyed.
+    accept_inplace: bool
+        Whether to allow inplace operations to already be present in the graph.
+
+    Raises
+    ------
+    TypeError
+        If inplace operations are not allowed and the graph already contains inplace operations.
+
+    """
+
+    has_destroy_handler = hasattr(fgraph, "destroyers")
+    if not (has_destroy_handler and accept_inplace):
+        # Check if fgraph already contains destructive operations,
+        # in which case we need to add a DestroyHandler or raise an error
+        for node in fgraph.apply_nodes:
+            if node.op.destroy_map:
+                if not accept_inplace:
+                    raise TypeError(
+                        f"Graph must not contain inplace operations: {node}"
+                    )
+                else:
+                    has_destroy_handler = True
+                    fgraph.attach_feature(DestroyHandler())
+                    break
+
+    # Protect all immutable inputs from inplace operations.
+    fgraph.attach_feature(
+        Supervisor(
+            input
+            for spec, input in zip(input_specs, fgraph.inputs, strict=True)
+            if not (
+                spec.mutable or has_destroy_handler and fgraph.has_destroyers([input])
+            )
+        )
+    )
+
+
 def std_fgraph(
     input_specs: list[SymbolicInput],
     output_specs: list[SymbolicOutput],
@@ -229,24 +283,8 @@
 
         found_updates.extend(map(SymbolicOutput, updates))
 
-    for node in fgraph.apply_nodes:
-        if node.op.destroy_map:
-            if not accept_inplace:
-                raise TypeError(f"Graph must not contain inplace operations: {node}")
-            else:
-                fgraph.attach_feature(DestroyHandler())
-                break
-
-    # We need to protect all immutable inputs from inplace operations.
-    fgraph.attach_feature(
-        Supervisor(
-            input
-            for spec, input in zip(input_specs, fgraph.inputs, strict=True)
-            if not (
-                spec.mutable
-                or (hasattr(fgraph, "destroyers") and fgraph.has_destroyers([input]))
-            )
-        )
+    add_supervisor_to_fgraph(
+        fgraph=fgraph, input_specs=input_specs, accept_inplace=accept_inplace
     )
 
     # If named nodes are replaced, keep the name

pytensor/compile/mode.py

@@ -138,7 +138,11 @@ def apply(self, fgraph):
                 break
         if not supervisor_added:
             warnings.warn(
-                f"A Supervisor feature is missing from {fgraph}.",
+                (
+                    f"A Supervisor feature is missing from {fgraph}.\n"
+                    "This is needed for inplace rewrites. Either exclude inplace rewrites or add a Supervisor feature.\n"
+                    "A Supervisor feature can be added via `pytensor.compile.function.types.add_supervisor_to_fgraph`."
+                ),
                 stacklevel=3,
             )
 

pytensor/link/jax/dispatch/scan.py

@@ -1,7 +1,7 @@
 import jax
 import jax.numpy as jnp
 
-from pytensor.compile.mode import JAX
+from pytensor.compile.mode import JAX, get_mode
 from pytensor.link.jax.dispatch.basic import jax_funcify
 from pytensor.scan.op import Scan
 
@@ -19,7 +19,9 @@ def jax_funcify_Scan(op: Scan, **kwargs):
         )
 
     # Optimize inner graph (exclude any defalut rewrites that are incompatible with JAX mode)
-    rewriter = op.mode_instance.excluding(*JAX._optimizer.exclude).optimizer
+    rewriter = (
+        get_mode(op.mode).including("jax").excluding(*JAX._optimizer.exclude).optimizer
+    )
     rewriter(op.fgraph)
     scan_inner_func = jax_funcify(op.fgraph, **kwargs)
 

pytensor/link/numba/dispatch/basic.py

@@ -16,9 +16,10 @@
 from numba.cpython.unsafe.tuple import tuple_setitem  # noqa: F401
 from numba.extending import box, overload
 
-from pytensor import config
+from pytensor import In, config
 from pytensor.compile import NUMBA
 from pytensor.compile.builders import OpFromGraph
+from pytensor.compile.function.types import add_supervisor_to_fgraph
 from pytensor.compile.ops import DeepCopyOp
 from pytensor.graph.basic import Apply
 from pytensor.graph.fg import FunctionGraph
@@ -430,7 +431,13 @@ def numba_funcify_OpFromGraph(op, node=None, **kwargs):
     # TODO: Not sure this is the right place to do this, should we have a rewrite that
     #  explicitly triggers the optimization of the inner graphs of OpFromGraph?
     #  The C-code defers it to the make_thunk phase
-    NUMBA.optimizer(op.fgraph)
+    fgraph = op.fgraph
+    add_supervisor_to_fgraph(
+        fgraph=fgraph,
+        input_specs=[In(x, borrow=True, mutable=False) for x in fgraph.inputs],
+        accept_inplace=True,
+    )
+    NUMBA.optimizer(fgraph)
     fgraph_fn = numba_njit(numba_funcify(op.fgraph, **kwargs))
 
     if len(op.fgraph.outputs) == 1:

pytensor/link/numba/dispatch/scan.py

@@ -4,7 +4,9 @@
 from numba import types
 from numba.extending import overload
 
-from pytensor.compile.mode import NUMBA
+from pytensor import In
+from pytensor.compile.function.types import add_supervisor_to_fgraph
+from pytensor.compile.mode import NUMBA, get_mode
 from pytensor.link.numba.dispatch import basic as numba_basic
 from pytensor.link.numba.dispatch.basic import (
     create_arg_string,
@@ -59,11 +61,18 @@ def numba_funcify_Scan(op, node, **kwargs):
     #  explicitly triggers the optimization of the inner graphs of Scan?
     #  The C-code defers it to the make_thunk phase
     rewriter = (
-        op.mode_instance.including("numba")
+        get_mode(op.mode)
+        .including("numba")
         .excluding(*NUMBA._optimizer.exclude)
         .optimizer
     )
-    rewriter(op.fgraph)
+    fgraph = op.fgraph
+    add_supervisor_to_fgraph(
+        fgraph=fgraph,
+        input_specs=[In(x, borrow=True, mutable=False) for x in fgraph.inputs],
+        accept_inplace=True,
+    )
+    rewriter(fgraph)
 
     scan_inner_func = numba_basic.numba_njit(numba_funcify(op.fgraph))
 

pytensor/link/numba/dispatch/subtensor.py

@@ -77,10 +77,10 @@ def convert_indices(indices, entry):
             y_name = input_names[1]
 
         if op.set_instead_of_inc:
-            function_name = "setsubtensor"
+            function_name = "set_subtensor"
             index_body = f"z[indices] = {y_name}"
         else:
-            function_name = "incsubtensor"
+            function_name = "inc_subtensor"
             index_body = f"z[indices] += {y_name}"
     else:
         function_name = "subtensor"

pytensor/link/pytorch/dispatch/basic.py

@@ -5,8 +5,10 @@
 import torch
 import torch.compiler
 
+from pytensor import In
 from pytensor.compile import PYTORCH
 from pytensor.compile.builders import OpFromGraph
+from pytensor.compile.function.types import add_supervisor_to_fgraph
 from pytensor.compile.ops import DeepCopyOp
 from pytensor.graph.basic import Constant
 from pytensor.graph.fg import FunctionGraph
@@ -185,6 +187,13 @@ def pytorch_funcify_OpFromGraph(op, node, **kwargs):
     kwargs.pop("storage_map", None)
     # Apply inner rewrites
     PYTORCH.optimizer(op.fgraph)
+    fgraph = op.fgraph
+    add_supervisor_to_fgraph(
+        fgraph=fgraph,
+        input_specs=[In(x, borrow=True, mutable=False) for x in fgraph.inputs],
+        accept_inplace=True,
+    )
+    PYTORCH.optimizer(fgraph)
     fgraph_fn = pytorch_funcify(op.fgraph, **kwargs, squeeze_output=True)
     return fgraph_fn
 

pytensor/scan/op.py

@@ -57,8 +57,9 @@
 from pytensor import tensor as pt
 from pytensor.compile.builders import construct_nominal_fgraph, infer_shape
 from pytensor.compile.function.pfunc import pfunc
+from pytensor.compile.function.types import add_supervisor_to_fgraph
 from pytensor.compile.io import In, Out
-from pytensor.compile.mode import Mode, get_default_mode, get_mode
+from pytensor.compile.mode import Mode, get_mode
 from pytensor.compile.profiling import register_profiler_printer
 from pytensor.configdefaults import config
 from pytensor.gradient import DisconnectedType, NullType, Rop, grad, grad_undefined
@@ -761,18 +762,7 @@ def __init__(
         self.profile = profile
         self.allow_gc = allow_gc
         self.strict = strict
-
-        # Clone mode_instance, altering "allow_gc" for the linker,
-        # and adding a message if we profile
-        if self.name:
-            message = f"{self.name} sub profile"
-        else:
-            message = "Scan sub profile"
-
-        self.mode = get_default_mode() if mode is None else mode
-        self.mode_instance = get_mode(self.mode).clone(
-            link_kwargs=dict(allow_gc=self.allow_gc), message=message
-        )
+        self.mode = mode
 
         # build a list of output types for any Apply node using this op.
         self.output_types = []
@@ -845,8 +835,6 @@ def tensorConstructor(shape, dtype):
         self.n_outer_inputs = info.n_outer_inputs
         self.n_outer_outputs = info.n_outer_outputs
 
-        _ = self.prepare_fgraph(self.fgraph)
-
         if any(node.op.destroy_map for node in self.fgraph.apply_nodes):
             raise InconsistencyError(
                 "Inner-graphs must not contain in-place operations."
@@ -1405,23 +1393,8 @@ def prepare_fgraph(self, fgraph):
 
         fgraph.update_mapping = update_mapping
 
-        from pytensor.compile.function.types import Supervisor
-        from pytensor.graph.destroyhandler import DestroyHandler
-
-        for node in fgraph.apply_nodes:
-            if node.op.destroy_map:
-                fgraph.attach_feature(DestroyHandler())
-                break
-
-        fgraph.attach_feature(
-            Supervisor(
-                inp
-                for spec, inp in zip(wrapped_inputs, fgraph.inputs, strict=True)
-                if not (
-                    getattr(spec, "mutable", None)
-                    or (hasattr(fgraph, "destroyers") and fgraph.has_destroyers([inp]))
-                )
-            )
+        add_supervisor_to_fgraph(
+            fgraph=fgraph, input_specs=wrapped_inputs, accept_inplace=True
         )
 
         return wrapped_inputs, wrapped_outputs
@@ -1445,10 +1418,17 @@ def fn(self):
         elif self.profile:
             profile = self.profile
 
+        # Clone mode_instance, altering "allow_gc" for the linker,
+        # and adding a message if we profile
+        mode_instance = get_mode(self.mode).clone(
+            link_kwargs=dict(allow_gc=self.allow_gc),
+            message=f"{self.name or 'Scan'} sub profile",
+        )
+
         self._fn = pfunc(
             wrapped_inputs,
             wrapped_outputs,
-            mode=self.mode_instance,
+            mode=mode_instance,
             accept_inplace=False,
             profile=profile,
             on_unused_input="ignore",

pytensor/sparse/rewriting.py

@@ -210,6 +210,7 @@ def local_inplace_addsd_ccode(fgraph, node):
     ),
     "fast_run",
     "inplace",
+    "cxx_only",
     position=50.1,
 )
 
@@ -241,6 +242,7 @@ def local_addsd_ccode(fgraph, node):
     WalkingGraphRewriter(local_addsd_ccode),
     # Must be after local_inplace_addsd_ccode at 70.0
     "fast_run",
+    "cxx_only",
     position=70.1,
 )
 

tests/link/numba/test_basic.py

@@ -835,6 +835,39 @@ def test_OpFromGraph():
     compare_numba_and_py([x, y, z], [out], [xv, yv, zv])
 
 
+@pytest.mark.filterwarnings("error")
+def test_ofg_inner_inplace():
+    x = pt.vector("x")
+    set0 = x[0].set(1)  # SetSubtensor should not inplace on x
+    exp_x = pt.exp(x)
+    set1 = exp_x[0].set(1)  # SetSubtensor should inplace on exp_x
+    ofg0 = OpFromGraph([x], [set0])
+    ofg1 = OpFromGraph([x], [set1])
+
+    y, z = pt.vectors("y", "z")
+    fn = function([y, z], [ofg0(y), ofg1(z)], mode="NUMBA")
+
+    fn_ofg0 = fn.maker.fgraph.outputs[0].owner.op
+    assert isinstance(fn_ofg0, OpFromGraph)
+    fn_set0 = fn_ofg0.fgraph.outputs[0]
+    assert fn_set0.owner.op.destroy_map == {}
+
+    fn_ofg1 = fn.maker.fgraph.outputs[1].owner.op
+    assert isinstance(fn_ofg1, OpFromGraph)
+    fn_set1 = fn_ofg1.fgraph.outputs[0]
+    assert fn_set1.owner.op.destroy_map == {0: [0]}
+
+    x_test = np.array([0, 1, 1], dtype=config.floatX)
+    y_test = np.array([0, 1, 1], dtype=config.floatX)
+    res0, res1 = fn(x_test, y_test)
+    # Check inputs were not mutated
+    np.testing.assert_allclose(x_test, [0, 1, 1])
+    np.testing.assert_allclose(y_test, [0, 1, 1])
+    # Check outputs are correct
+    np.testing.assert_allclose(res0, [1, 1, 1])
+    np.testing.assert_allclose(res1, [1, np.e, np.e])
+
+
 @pytest.mark.filterwarnings("error")
 def test_cache_warning_suppressed():
     x = pt.vector("x", shape=(5,), dtype="float64")