reference-generation

To generate a PyTorch nn.Module from a JSON operator description, use this Skill to parse the JSON and output a standalone Model class with proper type hints. The resulting file serves as a golden standard for testing and benchmarking.

Creating get_inputs and get_init_inputs helpers for PyTorch benchmarking is handled automatically during reference code generation. The Skill structures these functions alongside the main nn.Module to provide standardized test inputs directly from the operator JSON.

Yes, you can export generated PyTorch reference implementations to a defined output path. The Skill supports saving the generated Python code predictably, such as to the output/{op_name}_reference.py file, for easy workflow integration.

Yes, the generated PyTorch reference code strictly follows a Model(nn.Module) structure. This ensures the output is a self-contained, properly typed module that can be directly imported for regression testing or used as a design guide.

Using a JSON operator description to create a PyTorch reference implementation establishes a portable golden standard for documentation and testing. It converts abstract specifications into executable code, ensuring consistent validation for downstream implementations.