TorchLean API

Docs HomeInstallationGuideExamplesGraphsTools

NN.Runtime.PyTorch.Wire

Graph wire format #

NN.IR.OpTag identifies semantic operators. This module assigns each identity its fixed spelling in torchlean.ir.v1: for example, .mulElem is written as "mul_elem". Keeping this table separate from diagnostic names lets us improve an error message while keeping existing graphs readable. Exporters and importers use the same table.

The tag round trip is proved for every constructor. Static attributes are parsed separately; parseOpKind? constructs only operations that need none. Tuple projections and supported container nodes belong to the external value graph, not the tensor graph's semantic operator vocabulary.

source
def Interop.PyTorch.Wire.format :
String

Format marker written to the root format field.

Instances For
    source
    def Interop.PyTorch.Wire.tupleGetItem :
    String

    Projection of one component out of a tuple-valued FX node.

    Instances For
      source
      def Interop.PyTorch.Wire.mhaTuple :
      String

      Container-valued nn.MultiheadAttention call kept in the value graph.

      Instances For
        source
        def Interop.PyTorch.Wire.pyTuple :
        String

        Legacy marker for a tuple producer without a lowering rule; the importer reports its limits.

        Instances For
          source
          def Interop.PyTorch.Wire.opTag :
          NN.IR.OpTag → String

          Fixed constructor spelling in the torchlean.ir.v1 artifact.

          Instances For
            source
            def Interop.PyTorch.Wire.parseOpTag? (s : String) :
            Option NN.IR.OpTag

            Parse a v1 constructor spelling into its semantic identity.

            Instances For
              source
              def Interop.PyTorch.Wire.parseOpKind? (s : String) :
              Option NN.IR.OpKind

              Parse an attribute-free operation; attributed operators need their separate fields.

              Instances For
                source
                def Interop.PyTorch.Wire.kindField (tag : NN.IR.OpTag) :
                String

                JSON or Python "kind": "<wire>" field for an operator identity.

                Instances For
                  source
                  def Interop.PyTorch.Wire.kindObject (tag : NN.IR.OpTag) :
                  String

                  JSON or Python object holding only the kind field.

                  Instances For
                    source
                    def Interop.PyTorch.Wire.quotedOpTag (tag : NN.IR.OpTag) :
                    String

                    The wire string as a quoted JSON or Python literal.

                    Instances For
                      source
                      def Interop.PyTorch.Wire.quotedOpTags (tags : List NN.IR.OpTag) :
                      String

                      Python set literal of quoted wire strings.

                      Instances For
                        source
                        theorem Interop.PyTorch.Wire.parse_op_tag (tag : NN.IR.OpTag) :
                        parseOpTag? (opTag tag) = some tag

                        Every v1 tag parses back to its identity: the codec is complete and collision free.

                        source
                        theorem Interop.PyTorch.Wire.parse_op_kind_tag (kind : NN.IR.OpKind) :
                        parseOpTag? (opTag kind.opTag) = some kind.opTag

                        Serializing and parsing an operation preserves its identity, regardless of its attributes.

                        source
                        theorem Interop.PyTorch.Wire.parse_op_kind (kind : NN.IR.OpKind) (h : kind.opTag.hasAttributes = false) :
                        parseOpKind? (opTag kind.opTag) = some kind

                        The attribute-free v1 round trip reconstructs the original operation.