PyTorch Adapter¶

The vnvspec.torch module provides adapters that wrap PyTorch and HuggingFace models, implementing the ModelAdapter protocol for automated V&V assessment.

Installation¶

pip install vnvspec[torch]

This installs torch>=2.3 and transformers>=4.40 as additional dependencies.

TorchAdapter¶

TorchAdapter wraps any torch.nn.Module and runs it through a spec's requirements, producing an assessment Report.

from vnvspec.torch import TorchAdapter

adapter = TorchAdapter(
    model,
    device="cuda",
    sample_budget=1000,
    batch_size=32,
)
report = adapter.assess(spec, data_loader)
print(report.verdict())

Constructor parameters:

Parameter	Type	Description
`model`	`nn.Module`	The PyTorch model to assess
`input_adapter`	`Any`	Optional callable to preprocess inputs
`output_adapter`	`Any`	Optional callable to postprocess outputs
`device`	`str`	Device to run on (`"cpu"`, `"cuda"`)
`sample_budget`	`int`	Maximum number of samples to evaluate
`batch_size`	`int`	Batch size for evaluation (default 32)

Specialized Adapters¶

TransformerAdapter -- wraps HuggingFace transformer models with tokenizer integration
AutoregressiveAdapter -- handles autoregressive generation with stopping criteria
VLMAdapter -- wraps vision-language models with image preprocessing

Supporting Utilities¶

HookManager -- attach forward/backward hooks to model layers for intermediate activation inspection
SampleBudgetIterator -- wraps a DataLoader to enforce a maximum sample budget, stopping iteration once the budget is reached

ModelAdapter Protocol¶

All adapters implement the ModelAdapter protocol defined in vnvspec.core.protocols. Custom adapters for other frameworks (scikit-learn, ONNX, etc.) can implement this same protocol.

API reference: vnvspec.torch.adapter.TorchAdapter, vnvspec.torch.hooks.HookManager, vnvspec.torch.sampling.SampleBudgetIterator