| # Core ML is not shipped for this model | |
| Core ML cannot export this model at all, so no `.pte` or `config.json` is | |
| published here. Use the XNNPACK or MLX build instead. | |
| ## Why | |
| The exporter exists in `export-scripts` for all four sub-models, but conversion | |
| fails on the first one (`duration_predictor`), tested 2026-08-20: | |
| ``` | |
| Placeholder val must be a tensor or fake tensor, but got type | |
| <class 'torch.SymInt'>, value s16 + 1 | |
| ``` | |
| That is the Core ML dynamic-shape limitation hitting at **export** time rather | |
| than at execute. Every sub-model here carries dynamic dims (`T` for text 8..256, | |
| `L` for latent 4..1000), and a symbolic dim reaches the Core ML converter as a | |
| placeholder it cannot represent. | |
| `vocoder` takes only a float latent and was the best candidate of the four, but | |
| it is never reached: the run aborts on the first failure. | |
| Related: apple/coremltools#2825. | |