0.4.3:
All models now ship as our own exports, made from NVIDIA's NeMo checkpoints with PyTorch's dynamo exporter at ONNX opset 23 (one Attention op per layer, much smaller graphs) based on the reccomendation from onnxruntime/onnxruntime-ep-mlx#10 (comment). TDT, CTC, Unified and Multitalker now have their own folders in altunenes/parakeet-rs on Hugging Face, and every model has an export script in scripts.
Outputs match the previous exports on ONNX Runtime and speed is unchanged. The burn backend was regenerated from the new files, so burn users need to download the models again; ONNX Runtime users can keep their old files. However, my recommendation is that download the new versions to make sure compatibility with future updates. I haven't noticed any diffs in speed || performance in my own benchmarks.
What's Changed
- Move all ONNX exports to the dynamo exporter (opset 23) by @altunenes in #144
Full Changelog: v0.4.2...v0.4.3