g^(h ddlZddlmZddlmZddlmZddlZdjZ GddejjZ dejd ejjd ejjfd Zd eed fdej&dej(ded ejjf dZej.Zed eeddffdZd efdZy)N) Generator)contextmanager)Anya This is used when dynamo traces torch.nn.Parameter, which normally would not trace properly with AOTAutograd. We instead create a placeholder torch.nn.Parameter before the graph, which becomes a graph arg and has no storage backing it. At the point in the graph where the parameter actually should be created we mutate this sacrificial placeholder into it. This allows gradients to flow into the parameter as if it were an input to the graph (which is the only thing we are allowed to compute gradients on). c eZdZededededej jfdZededejde dejffdZ y) TracableCreateParameterctxtensor placeholderreturnc@|jrJ|j|SN) requires_gradset_)rr r s _/var/www/skyplay_api_hub/venv/lib/python3.12/site-packages/torch/_dynamo/create_parameter_op.pyforwardzTracableCreateParameter.forwards"'''''' grad_outputsNc|d}d|fS)Nr)rrgrads rbackwardz TracableCreateParameter.backwardsATzr) __name__ __module__ __qualname__ staticmethodrtorchnn ParameterrTensortuplerrrrrrsu(S(#(C(EHHrBs%%  EG enn55  LL',xx'9'9 XX  S/ "'++ 7<|| TX  XX y3idD8H.I33=4=r