tf2onnx) → torch
(onnx2torch + the custom op shim in convert_ops.py) and traced to TorchScript.text_encoder.pt — text tower: (ids[1,128] int32, paddings[1,128] f32) → [1,768]quantizer.pt — RVQ: [1,768] → [12,1] style tokensspm.model — SentencePiece vocabtorch + sentencepiece only. Use via magenta_rt.torch.musiccoca.MusicCoCa.