MCPcopy Create free account
hub / github.com/pytorch/executorch / build_graph

Function build_graph

backends/qualcomm/utils/utils.py:812–867  ·  view source on GitHub ↗
(
        inputs,
        outputs,
        qnn_in_order: Optional[List[int]] = None,
        executorch_in_order: Optional[List[int]] = None,
        executorch_out_order: Optional[List[int]] = None,
    )

Source from the content-addressed store, hash-verified

810 )
811
812 def build_graph(
813 inputs,
814 outputs,
815 qnn_in_order: Optional[List[int]] = None,
816 executorch_in_order: Optional[List[int]] = None,
817 executorch_out_order: Optional[List[int]] = None,
818 ):
819 # custom op declaration
820 inputs_str = "Tensor[] inputs"
821 func_proto = f"{op_name}({inputs_str}) -> Any"
822 custom_op = Library(OpContextLoader.namespace, "FRAGMENT")
823 custom_op.define(func_proto)
824 # custom op implementation
825 implement_op(custom_op, op_name, outputs)
826
827 # model architecture mimicking context binary
828 class Model(torch.nn.Module):
829 """
830 The args of forward() can be thought of as what executorch is accepting as input.
831 The getattr inside the forward() can be thought of as qnn context binary.
832 When we first pass in the input, we need to use the executorch's(nn.module) input order.
833 After we get into forward(), we then need to convert input order to qnn's input order.
834 Same as return, when qnn returns the value, we need to reorder them back to executorh's output order.
835 """
836
837 def __init__(self, qnn_in_order, executorch_out_order):
838 super().__init__()
839 self.qnn_in_order = qnn_in_order
840 self.executorch_out_order = executorch_out_order
841
842 def forward(self, *inputs): # executorch
843 if self.qnn_in_order:
844 inputs = tuple(inputs[i] for i in self.qnn_in_order)
845 ret = getattr(
846 getattr(torch.ops, OpContextLoader.namespace), op_name
847 ).default(inputs)
848 return (
849 [ret[idx] for idx in self.executorch_out_order]
850 if self.executorch_out_order
851 else ret
852 )
853
854 inputs = (
855 tuple(tuple(inputs.values())[i] for i in executorch_in_order)
856 if executorch_in_order
857 else tuple(inputs.values())
858 )
859
860 model = Model(qnn_in_order, executorch_out_order)
861 prog = torch.export.export(model, inputs, strict=True)
862 # bookkeeping for variables' life cycle
863 return {
864 "custom_op": custom_op,
865 "custom_module": model,
866 "exported_program": prog,
867 }
868
869 def build_tensor(tensors, dtype_map):

Callers 1

from_context_binaryFunction · 0.85

Calls 4

implement_opFunction · 0.85
valuesMethod · 0.80
ModelClass · 0.70
exportMethod · 0.45

Tested by

no test coverage detected