task-constructor

Creating a self-contained KernelBench task from PyTorch code requires extracting operator implementations, tracing AST dependencies, inlining functions, and cleaning imports to form a validated single-file task. This skill automates that packaging process for reproducible benchmarking.

Inlining dependencies and cleaning imports for Triton codebases ensures the final task file is completely self-contained, removing external references so the code can run independently for reproducible KernelBench workflows and reference testing.

Yes, packaging code fragments for reference testing in KernelBench is supported. The tool extracts the target code, inlines required functions, validates the format, and enables direct reference comparisons through the provided testing tooling.

PyTorch must be installed to convert code into a KernelBench task. The dependency is required because the skill extracts operator implementations directly from PyTorch codebases and validates them against reference testing workflows.

The best way to extract PyTorch operator implementations for reproducible benchmarking is using AST dependency tracking to identify required functions, inline them, and clean imports to generate a standardized single-file task.

Extracted PyTorch code fails KernelBench format validation when external dependencies are not inlined or imports are not cleaned. The task must be a fully self-contained single file to pass validation and execute reproducible benchmarking.