+ j ^RIt^RIt]PPRR4t]P R4]PP ]RR4]PP ]RR4RR l44tR tR#) NexportFRAGMENTzLaccess_subclass_inner_tensor(Tensor src_subclass_tensor, str attr) -> Tensoraccess_subclass_inner_tensorAutogradPythoncdV^8dQhR\PR\R\P/#)src_subclass_tensorattrreturn)torchTensorstr)formats"X/home/luca/screen-checking/.venv/lib/python3.14/site-packages/torch/export/custom_ops.py __annotate__rs,-0 \\c^RIHpV!V4'g\R\V4 24h\ WR4pVe!\ V\ P4'g\RV RV 24hV#)r)is_traceable_wrapper_subclasszIExpected src_subclass_tensor to be a traceable wrapper subclass, but got Nz Attribute z% is not a tensor or doesn't exist in ) torch.utils._python_dispatchrAssertionErrortypegetattr isinstancer r RuntimeError)r r rvals&& r_access_subclass_inner_tensorrsK ()< = =/01 3   %T 2C {*S%,,77CDWCX Y   JrcVPR^4wr4\P!V4p\WT4p\ VR4'g\ RV R24hVP !V/VB#)aA Import a custom autograd function by string name and call it. This is pretty bad because: 1) There is no schema Ideally we should automatically wrap custom autograd functions with a custom op, but that is too much work because we need to schematize custom autograd functions. For now, we just hackily put it in the IR. .applyzExpected function class z to have 'apply' method)rsplit importlib import_modulerhasattrrr )function_cls_nameargskwargs module_name class_namemodule function_clss&*, r._call_custom_autograd_function_in_pre_dispatchr,$sv066sA>K $ $[ 1F6.L < ) )&'8&99P Q     t .v ..r) r"r libraryLibrarylibdefineimplrr,rrr3s  mmHj1 R C7DC7BCE$/r