Module solbench 

SOL-ExecBench harness bridge — turn a recognized LOGOS fused-row op into a submittable solution.json plus a torch-binding CUDA kernel that conforms to the benchmark’s ABI.

The benchmark evaluates sol-execbench <problem_dir> --solution solution.json. A CUDA/C++ solution is a JSON manifest referencing source files, with an entry_point "file::fn". The entry point is a Destination-Passing-Style binding — void run(torch::Tensor inputs…, torch::Tensor output) — that takes the inputs (in Definition order) by value and writes the pre-allocated output in place. The harness materializes the reference run() (PyTorch) and compares within (atol, rtol, matched_ratio) on a B200.

Our emitters produce a self-contained checksum program for the ≡-oracle GPU locks; this bridge wraps the SAME kernel math in the harness ABI instead: a runtime-shaped kernel (rows/cols are kernel args, not baked) and a run binding that up-casts the (bf16/…) inputs to f32, launches, and writes the output back.

Structs

SolSolution

A generated SOL-ExecBench solution: the manifest and the CUDA source it references.

Functions

emit_fused_row_solution

Emit a SOL-ExecBench solution for a recognized fused-row op (RMSNorm plain/weighted/fused-add, or LayerNorm). definition is the problem’s Definition name; inputs the input tensor names in Definition order; output the output name. The emitted run binding matches that order (DPS).