Merge: cached_property pytree/dict leak structural defense

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: 2 people

autoarray/abstract_ndarray.py

@@ -102,8 +102,14 @@ def register_instance_pytree(cls, no_flatten=()):
     if cls in _pytree_registered_classes:
         return
     from autoconf.jax_wrapper import register_pytree_node
+    from autoconf.tools.decorators import cached_property_names
 
-    no_flatten_set = frozenset(no_flatten)
+    # Extend the caller-supplied no_flatten set with every
+    # ``cached_property``-style descriptor on ``cls`` so derived caches
+    # (e.g. heavy plotting/aggregator @cached_property values) never reach
+    # the pytree leaves and break ``jax.jit``. Mirrors the PyAutoFit-side
+    # defense; see PyAutoFit#1300 for the diagnosed class of bug.
+    no_flatten_set = frozenset(no_flatten) | cached_property_names(cls)
 
     def flatten(instance):
         dyn: list = []
@@ -174,12 +180,19 @@ def instance_flatten(cls, instance):
         """
         Flatten an instance of an autoarray class into a tuple of its attributes (i.e.. a pytree)
         """
+        from autoconf.tools.decorators import cached_property_names
+
+        # Union the class-level ``__no_flatten__`` opt-out with any
+        # ``cached_property`` descriptor names so derived caches don't
+        # surface as pytree leaves. Mirrors the defense in
+        # ``_register_as_pytree`` (PyAutoFit#1300 follow-up).
+        excluded = set(cls.__no_flatten__) | cached_property_names(cls)
         keys, values = zip(
             *sorted(
                 {
                     key: value
                     for key, value in instance.__dict__.items()
-                    if key not in cls.__no_flatten__
+                    if key not in excluded
                 }.items()
             )
         )

test_autoarray/test_abstract_ndarray_pytree_guard.py

@@ -0,0 +1,82 @@
+"""Regression: cached_property descriptors on AbstractNDArray subclasses
+must be filtered from ``instance_flatten`` so derived caches never reach
+the JAX pytree leaves.
+
+NumPy-only per the project rule [[feedback_no_jax_in_unit_tests]]:
+exercise the ``instance_flatten`` classmethod directly (which is what
+the JAX pytree path delegates to) and assert composition is correct.
+"""
+
+import functools
+
+import numpy as np
+
+from autoarray.abstract_ndarray import AbstractNDArray
+
+
+class _FakeArray(AbstractNDArray):
+    """Minimal AbstractNDArray subclass that adds a ``@cached_property``
+    returning a string. Used to assert the guard filters it from
+    ``instance_flatten``."""
+
+    __no_flatten__ = ("use_jax",)
+
+    def __init__(self, array):
+        # Skip AbstractNDArray.__init__ to avoid the JAX-registration path
+        # — we only need the dict-shape for the flatten test.
+        self._array = np.asarray(array)
+        self._is_transformed = False
+        self.use_jax = False
+
+    @property
+    def native(self):
+        # AbstractNDArray declares ``native`` abstract; the body is
+        # irrelevant to the flatten path so just echo ``_array``.
+        return self._array
+
+    @functools.cached_property
+    def heavy_summary(self):
+        return "a-pretty-printed-summary-of-the-array"
+
+
+def test_instance_flatten_excludes_cached_property_names():
+    """``AbstractNDArray.instance_flatten`` unions the class-level
+    ``__no_flatten__`` with the result of
+    ``autoconf.tools.decorators.cached_property_names`` so derived
+    cached strings stay out of the pytree leaves.
+
+    This pins the structural defense that follows PyAutoFit#1300: the
+    leak surfaces today only on the Model side, but the same opt-out
+    filter shape on AbstractNDArray descendants would break ``jax.jit``
+    the moment anyone added a ``@cached_property`` returning a
+    non-array value to a Fit class."""
+
+    arr = _FakeArray([1.0, 2.0, 3.0])
+
+    # Trigger the cached property: it writes "...summary..." into __dict__.
+    _ = arr.heavy_summary
+    assert arr.__dict__["heavy_summary"] == "a-pretty-printed-summary-of-the-array"
+
+    leaves, keys = _FakeArray.instance_flatten(arr)
+
+    # The pre-existing __no_flatten__ exclusion ("use_jax") still applies.
+    assert "use_jax" not in keys
+    # The new cached_property exclusion fires too.
+    assert "heavy_summary" not in keys
+    # No string leaves anywhere.
+    assert not any(isinstance(leaf, str) for leaf in leaves)
+
+
+def test_instance_flatten_preserves_array_data():
+    """Sanity check: filtering cached_property names does not collateral-
+    damage real array data. The underlying numpy array must still appear
+    in the leaves."""
+
+    arr = _FakeArray([1.0, 2.0, 3.0])
+    _ = arr.heavy_summary  # poison the cache before flattening
+
+    leaves, keys = _FakeArray.instance_flatten(arr)
+
+    assert "_array" in keys
+    array_index = keys.index("_array")
+    np.testing.assert_array_equal(leaves[array_index], np.asarray([1.0, 2.0, 3.0]))