PyTorch Round-Trip Driver

This module assembles the state-dict round-trip examples.

It does not re-implement model math. Instead it wires together:

• PyTorch exporters (MLP/Export, CNN/Export, Transformer/Export)
• typed PyTorch JSON importers (MLP/Import, CNN/Import, Transformer/Import)
• native tensor operations for the MLP forward pass
• public executable nn modules for the CNN and Transformer forward passes

Run via the TorchLean example runner:

scripts/lake.sh exe torchlean pytorch_roundtrip --model mlp --action import

Choose mlp, cnn, or transformer for --model, and export or import for --action.

Design goals:

• keep paths/dimensions centralized (no duplicated constants across examples),
• keep the example output deterministic and readable,
• keep this example import-safe (no root-level main that collides with other executables).

CLI model/action selection

Paths and fixed example dimensions

Small IO helpers

Export actions

Import actions

def NN.Examples.Interop.PyTorch.Roundtrip.mlpOutput : IO (TorchLean.Tensor Float [1])

Load the MLP JSON fixture and evaluate [0.5, 0.8] with native Float tensor operations.

def NN.Examples.Interop.PyTorch.Roundtrip.cnnOutput : IO (TorchLean.Tensor Float [2])

Load the CNN JSON fixture and evaluate the row-major input 1..64 on CPU with scalar Float.

def NN.Examples.Interop.PyTorch.Roundtrip.transformerOutput : IO (TorchLean.Tensor Float [1, 2])

Load the encoder JSON fixture and evaluate the token [[1.5, 1.5]] on CPU with scalar Float.

Public entrypoint called from the examples runner

def NN.Examples.Interop.PyTorch.Roundtrip.main (args : List String) : IO Unit

Run the selected import or export action; omitted flags select MLP export.