NN.Runtime.PyTorch.Import

Reusable PyTorch weight-import surface.

The general import path is JSON-first:

1. PyTorch loads the original checkpoint / state_dict.
2. A caller-supplied Python script writes the tensors as nested-list JSON.
3. Import.PyTorch parses that JSON into shape-checked TorchLean tensors.

For graphs, the matching path is:

1. PyTorch captures an nn.Module with the adapter emitted by Export.PyTorch.TorchExport.script.
2. Import.PyTorch.TorchExport.parseGraph reads the torchlean.ir.v2 JSON into NN.IR.Graph.
3. The parser runs the shared IR well-formedness and shape checkers before accepting the graph.

For per-node element types and explicit binary casts, use Import.IR.parseProgram from NN.Runtime.IR. Its exact-word payload loader supplies typed constants and linear parameters. The Python torchlean.verify entry point uses this path; other parameterized operators still need exact typed payload loaders before binary execution can run them.

Op kind strings in that artifact are resolved through Interop.PyTorch.Wire, the same table the exporters write with; parsing a serialized operator tag recovers its NN.IR.OpTag identity.

This is the supported graph-import path today. The ONNX adapter emitted by Export.PyTorch.ONNX.script follows the same rule: lower ONNX nodes into the torchlean.ir.v2 artifact, then reuse this parser and the shared IR validators. Keeping the artifact boundary explicit avoids a second, looser graph semantics.

Graph import and payload import remain separate phases. A Conv/Gemm/BatchNorm graph can be validated as TorchLean IR once the adapter expands it into supported nodes; executing the graph then requires the matching constants/linear/conv payload store.

Import.PyTorch.MLP, Import.PyTorch.CNN, and Import.PyTorch.Transformer provide model-family adapters. Their runnable examples and reference artifacts live under NN.Examples.Interop.PyTorch.