KernelOPT uses five profiling-guided agents to optimize only the Triton sub-kernels generated by a compiler while preserving cuBLAS and cuDNN dispatches. Candidates pass static validation, multi-seed correctness, model-level float64-fallback checks and performance gates before being reinserted. If none passes, the compiler baseline remains. On 250 KernelBench problems, geometric-mean speedups over torch.compile were 1.40x, 1.15x and 1.07x across the three difficulty levels. The repository named in the paper returned 404 during this run.
KernelOPT verifies agent-optimized GPU kernels end to end
2026-09-25Frontier watchWATCH
KernelOPT uses five profiling-guided agents to optimize only the Triton sub-kernels generated by a compiler while preserving cuBLAS and cuDNN dispatches. Candidates pass static validation, multi-seed correctness, model-level float64-fallback checks and performance gates before being reinserted. If none passes, the compiler baseline remains. On 250 KernelBench problems, geometric-mean speedups over torch.compile were 1.40x, 1.15x and 1.07x across the three difficulty levels. The repository named in the paper returned 404 during this run.
Imported from the existing research archive. Migration preserves the recorded judgment and is not a fresh verification.
What this means for you
For inference teams, the important idea is the acceptance layer: agent-generated low-level code should never ship on a single unit test or microbenchmark. Preserve proven library paths, verify the reassembled model, and keep the compiler output as the automatic fallback.
FRONTIER WATCH
Public artifacts and validated results are different.
The recorded fact is the public work and its artifacts. Author-reported outcomes do not establish independent reproduction. Next checks include usable code, task conditions and third-party validation.
Evidence & related material
Technical details & scores
| Search | five agents guided by NVIDIA Nsight Compute profiling |
|---|---|
| Scope | generated Triton sub-kernels; preserve cuBLAS and cuDNN dispatches |
| Verification | static, multi-seed, model-level float64 fallback and performance gates |
| Benchmark | 250 KernelBench problems |
| Speedup | 1.40x Level 1; 1.15x Level 2; 1.07x Level 3 geometric mean over torch.compile |
| Inputs | PyTorch nn.Module, Triton and Helion |
| Artifact | claimed repository returned 404 during this run |
Technical impact 4
Engineering value 4
Adoption signal 1
Maturity 2
Verification cost 4
Risk 2