PyTorch Dynamo adds BoundMethodVariable to fix recursive method dispatch

PyTorch's latest commit to its Dynamo JIT compiler introduces BoundMethodVariable, a new variable type designed to handle method dispatch on bound method objects. The patch addresses a critical infinite recursion bug that occurred when object_generic_getattr found a method with custom call_method logic. Previously, tp_descr_get_impl would call obj.getattro_impl, re-entering object_generic_getattr and looping indefinitely. Now, the system raises an _UnhandledDescriptorError instead; this error propagates to generic_getattr's NotImplementedError catch, falling back safely to GetAttrVariable.

The BoundMethodVariable models Python's method object with python_type, richcompare_impl, and hash_impl, ensuring accurate method dispatch. The call_function method now correctly routes through call_method on the bound object. Co-authored by Claude Code (Anthropic's AI assistant), the commit includes extensive test coverage: 47 OK in test_tp_getattro.py, 516 in test_functions.py, and 805 in test_misc.py. This bug fix is particularly valuable for production systems relying on PyTorch's eager-mode optimization, as it resolves a subtle source of crashes and performance regressions in method-heavy models.