| """SOTA reference for TopK: torch.topk itself. | |
| torch.topk dispatches to a CUB-backed kernel that uses radix-select for | |
| moderate k and a tuned bitonic sort for small n. It is the bar the agent's | |
| hand-rolled bitonic kernel must beat. There is no obvious vendor library that | |
| does better on the (batch, n, k) shape mix we evaluate — Faiss BlockSelect is | |
| specialized for k>=32 with much larger n, and CUB's DeviceSegmentedRadixSort | |
| sorts the full row (overkill for top-k). | |
| This file is INTENTIONALLY allowed to call torch.topk because it is the SOTA | |
| oracle, not the agent's submission. The agent's solution.py is forbidden from | |
| using torch.topk (see problem.yaml.forbidden). | |
| """ | |
| from __future__ import annotations | |
| import torch | |
| def sota_forward(x: torch.Tensor, k: int): | |
| """Best-available top-k reference. x: (batch, n) fp32.""" | |
| return torch.topk(x, k=k, dim=-1, largest=True, sorted=True) | |
| def is_available() -> bool: | |
| return True | |