task-constructor

To convert PyTorch code into KernelBench tasks, you provide the path to a PyTorch or Triton codebase. The skill extracts operator implementations and packages them into a single-file, self-contained task.

AST dependency tracking identifies and resolves required functions during code extraction. It ensures the generated KernelBench task remains portable and reproducible by automatically inlining dependencies and cleaning imports.

This tool works with both PyTorch and Triton codebases. It extracts operator implementations from either framework and packages them into self-contained KernelBench tasks with automated dependency inlining.

The best way to package self-contained operator implementations is using automated AST dependency tracking and import cleanup. This extracts the operator, inlines required functions, and formats it into a reproducible single file.

To validate generated KernelBench tasks, the skill performs format verification and reference-based testing. This ensures the extracted task maintains consistency with the original PyTorch or Triton implementation.

Extracted PyTorch tasks may fail format verification if dependency tracking or import cleanup is incomplete. The skill automates AST inlining and validation to enforce single-file format consistency and prevent these errors.