RAVE can be exported to ONNX format for deployment in environments that do not support TorchScript

In addition to TorchScript (.ts) export for nn~/VST use, RAVE provides an export_onnx script that serializes the model to the ONNX (Open Neural Network Exchange) format via torch.onnx.export. The onnx architecture config is a noiseless v1-equivalent configuration suitable for ONNX export (minimum 6 GB GPU). ONNX enables deployment in inference runtimes beyond PyTorch (e.g., ONNX Runtime, CoreML, TensorRT), which may offer better performance on specific hardware or integrate more easily with non-Python hosts. This path is less common for live music performance than the TorchScript/nn~ route.

Examples

rave train --config onnx ... then rave export_onnx --run /path/to/run. The resulting .onnx file can be loaded by any ONNX-compatible runtime.

Assessment

When would you choose ONNX export over TorchScript export for a RAVE model? What configuration is required during training to support ONNX export?