agents

Using agents to replace ONNX Runtime, cuDNN and cuBLAS with cuda-oxide kernels for a 32× speedup

On the same RTX 4090 model, speakrs now diarizes VoxConverse dev 32× faster and Earnings-21 36× faster than its old ONNX Runtime CUDA path, with the same DER. That's about 60× faster than pyannote on the same GPU, and an RTX 5090 runs at 2922× real time. AI agents wrote most of the kernels. Here is what we changed, what we measured, and what we'd do differently.