KernelOPT is a multi-agent system for GPU kernel optimization that works on compiled PyTorch models instead of standalone kernels. It keeps the compiler's vendor library calls (cuBLAS, cuDNN) and targets only the generated Triton sub-kernels. Five profiling-guided LLM agents propose candidate kernels, and a four-gate cascade filters them: static validation, multi-seed correctness, model-level float64-fallback , and performance gating. The re-stitched model is verified end-to-end; if nothing passes, the compiler baseline is kept.
Across 250 KernelBench problems, KernelOPT reports geometric mean speedups over torch.compile of 1.40x at Level 1, 1.15x at Level 2, and 1.07x at Level 3. This matters because prior LLM-based kernel optimizers often ignored the compiler's dispatch decisions and did not verify the whole model, so their gains were hard to trust in real inference and training paths.