Summary

Destroying a torch.library.Library (DEF mode) after its op has been called leaves torch.ops in a poisoned state: the _OpNamespace still lists the namespace and OpOverloadPacket._overload_names still lists 'default', but getattr(packet, 'default') raises AttributeError because the underlying op has been deregistered from the C++ dispatcher.

Any subsequent code that walks all CompositeImplicitAutograd ops crashes — most visibly, {**core_aten_decompositions()} at torch/_inductor/decomposition.py:108 surfaces the underlying AttributeError as the misleading TypeError: 'CustomDecompTable' object is not a mapping (CPython's DICT_UPDATE opcode rewrites AttributeError that way).

Repro

import torch
from torch.library import Library, impl

my_lib = Library('my_lib', 'DEF')
my_lib.define('my_func() -> None')

@impl(my_lib, 'my_func', '')
def my_func(): pass

torch.ops.my_lib.my_func()   # required
del my_lib                    # required

from torch._inductor.cudagraph_trees import reset_cudagraph_trees
# TypeError: 'CustomDecompTable' object is not a mapping

The actual error, hidden by the dict-spread, is at torch/_export/utils.py:1368 in _collect_all_valid_cia_ops_for_namespace:

AttributeError: The underlying op of 'my_lib.my_func' has no overload name 'default'

How this manifests in CI

slow-gradcheck periodic jobs (e.g. job 73398436183 in run 25055128899) hit this in test_torch.py after test_storage_preserve_nonhermetic_in_hermetic_context poisons the namespace. Subsequent CUDA tests then fail in before_cuda_memory_leak_check → torch._dynamo.reset() → lazy import chain into torch._inductor.decomposition.

The trigger landed in #181040, which added the before_cuda_memory_leak_check hook calling torch._dynamo.reset(). The latent bug in CIA op enumeration predates that — likely from #137650 (lazy decomp table) — but only became reachable mid-suite after #181040.

The flaky-test disabler has been auto-creating per-test "DISABLED ..." issues (#181343, #181684, #181537, ...) that paper over the symptom without addressing the root cause.

Possible fixes

1. _collect_all_valid_cia_ops_for_namespace skips overloads that no longer resolve (cheapest, most defensive).
2. Library.__del__ cleans up the corresponding _OpNamespace cache and torch.ops._dir entries (most correct, but trickier — multiple Library objects can share a namespace).
3. OpOverloadPacket.overloads() re-queries the dispatcher rather than returning the cached _overload_names.

Independently, torch/_inductor/decomposition.py:108 should arguably use dict(core_aten_decompositions()) instead of {**...} so the real error isn't masked.

cc @anjali411 @chauhang @penguinwu @bdhirsh @bobrenjc93 @jansel @tugsbayasgalan