op-task-extractor

To extract PyTorch operators into a KernelBench task, provide your PyTorch source code path and an optional shape/dtype source. The Skill automatically inlines dependencies and generates a self-contained task_desc.py file with Model.forward, get_inputs, and get_init_inputs functions.

The best way to convert PyTorch code into a single-file task_desc.py is using an automated extraction process that inlines dependencies. It wraps your operator implementation into a self-contained Model.forward() method while preserving deterministic execution for reproducible benchmarking.

To infer input and output signatures for KernelBench tasks, supply an additional shape and dtype information source alongside your PyTorch code. The extractor uses these sources to infer signatures and generate the required get_inputs() and get_init_inputs() methods.

Yes, you can validate generated KernelBench tasks for correctness using the built-in validation workflow. This workflow ensures the extracted PyTorch operator implementation is accurate by performing both static and runtime checks on the packaged single-file task.

Yes, the PyTorch operator extraction process supports external dependencies by automatically inlining them into the generated file. This ensures the final KernelBench-ready task_desc.py remains completely self-contained and executable without missing module errors.