#!/usr/bin/env python3 """Inspectable map to the full pinned CuTe DSL GEMM implementation. CuTe DSL GEMM is not honestly reducible to a few lines: the implementation must define layouts, tiled copies, MMA atoms, pipelines, synchronization, and a host launcher. The pinned CUTLASS source is the executable authority. This file keeps those boundaries visible without inventing a fake kernel. """ PINNED_SOURCE = ( "https://github.com/NVIDIA/cutlass/tree/v4.5.1/" "examples/python/CuTeDSL" ) COMPILATION_PATH = ( "Python @cute.jit host launcher", "tensor/layout construction", "@cute.kernel GPU function", "global-to-shared tiled copy", "shared-to-register tiled copy", "cute.gemm over an MMA atom", "epilogue store", "CuTe MLIR and NVIDIA lowering", "PTX and target machine code", ) def inspect() -> dict[str, object]: return { "source": PINNED_SOURCE, "path": COMPILATION_PATH, "coverage_state": "parser_only", "observation_state": "not_observed", "reason": "Full official kernel is pinned; no local GPU artifact exists.", } if __name__ == "__main__": print(inspect())