Pettorch
Pettorch connects PyTorch to MeTTa through this library's public integration surface. Arrays use metta.arrays, losses and optimizers use metta.integrate, and the package adds the behavior that is specific to PyTorch: autograd operations, module wrappers, training helpers, architecture reflection, and neural predicates.
Install from the checkout
Choose and install the PyTorch build for your CPU, CUDA, or ROCm environment with the official PyTorch selector. Then run this command from the pettorch checkout:
pip install .Pettorch's pyproject.toml defines no torch extra and no runtime dependencies. PyTorch is deliberately chosen by the user because the correct build depends on the machine. The package imports PyTorch only when a PyTorch-backed feature is first used.
Run the first tensor expression
In Python, import metta and pettorch, create m = metta.MeTTa(), and call pettorch.install(m). Then evaluate m.run("!(t-tolist (matmul (tensor ((1.0 2.0))) (tensor ((3.0) (4.0)))))"). The expression multiplies a one-by-two tensor by a two-by-one tensor and returns ((11.0)) as a MeTTa expression.
pettorch.install(m) installs the shared array operations with PyTorch as the constructor default, then adds the PyTorch-specific operations. Tensors cross the engine boundary as the same live objects, so their autograd graph remains attached.
Surface
| Surface | Direction |
|---|---|
install | Register tensor operations, losses, and model reflection. |
wrap | Expose a callable torch.nn.Module as a MeTTa function and reflect its structure. |
MettaModule | Wrap a MeTTa forward function as a torch.nn.Module with live parameters. |
neural_predicate | Register a network as a weighted relation over class atoms. |
reflect | Write module, child, parameter, shape, and linear-layer facts into a space. |
attach_optimizer and train_step | Expose optimizer effects or run one checked training step. |
EmbeddingStore | Re-export the shared array-backed nearest-neighbour store. |
A wrapped PyTorch module lets MeTTa rules choose which model runs. MettaModule goes the other way: its forward evaluates a named MeTTa function, and parameters are available to equations through (param name). The forward pass must answer exactly one tensor.
neural_predicate registers a network as an annotated relation: each class answers with its softmax probability as the answer's annotation, so (top 1 ...) selects the argmax and (annotation) reads each probability beside its class.
Reflection writes ordinary facts such as nn-module, nn-child, nn-param, nn-param-shape, and nn-linear. Rules can match model structure beside application facts.
See the shared metta.arrays, metta.integrate, and weighted relations pages for the MeTTa side of these integrations.
Repository status
The audited pettorch checkout at commit a501f6e has no configured remote, and no separate public pettorch repository is currently available to link. The public MeTTa repository describes pettorch as a sibling project. A pettorch URL should be added here only after that checkout has an actual public remote.