# SOL-ExecBench kernels

10 of the 235 GPU kernel implementations produced by Recursive's automated
AI-research system for the
[NVIDIA SOL-ExecBench](https://github.com/NVIDIA/SOL-ExecBench) leaderboard,
shared as illustrative examples. The rest stay private to avoid biasing the
leaderboard.

Each file is a single implementation of one benchmark kernel, scored on B200.
Files are named after the kernel they implement. Scores below are **SOL
score**, the standard metric on SOL-ExecBench (higher is better).

| Kernel | Category | SOL score |
|---|---|---|
| `011_fp8_moe_gate_routing.py` | Quant | 0.866 |
| `012_gqa_paged_decode_h32_kv4_d128_ps1.py` | FlashInfer-Bench | 0.934 |
| `025_moe_expert_parallel_execution_backward.py` | L2 | 0.533 |
| `032_nvfp4_moe_expert_linear_with_gating.py` | Quant | 0.906 |
| `045_fused_linear_gelu_grn_linear.cu` | L1 | 0.733 |
| `049_group_limited_topk_routing.cu` | L2 | 0.884 |
| `057_residual_coupling_flow_block.py` | L2 | 0.721 |
| `069_joint_transformer_block_residual_path.py` | L2 | 0.831 |
| `073_fused_encoder_final_norm_to_decoder_cross_attention_kv_projection.py` | L1 | 0.846 |
| `077_whisper_decoder_output_projection.py` | L1 | 0.720 |

Categories follow the SOL-ExecBench leaderboard taxonomy: L1 (single-op
kernels), L2 (multi-op fused blocks), Quant (FP8 / NVFP4), and FlashInfer-Bench
(paged-attention kernels measured against the FlashInfer reference).

## License

Apache-2.0. See the top-level [`LICENSE`](../LICENSE).