Conversation
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting
| def _resolve_bundle(args: argparse.Namespace) -> CoreMLModelBundle: | ||
| base = Path(args.model_dir) if args.model_dir else None | ||
| if base is None and not all([args.preprocessor, args.encoder, args.decoder, args.joint, args.tokenizer]): | ||
| raise ValueError("Either --model-dir or explicit model paths are required") | ||
| return CoreMLModelBundle( | ||
| preprocessor=Path(args.preprocessor) if args.preprocessor else base / "Melspectrogram.mlpackage", | ||
| encoder=Path(args.encoder) if args.encoder else base / "ParakeetEncoder.mlpackage", | ||
| decoder=Path(args.decoder) if args.decoder else base / "ParakeetDecoder.mlpackage", | ||
| joint=Path(args.joint) if args.joint else base / "RNNTJoint.mlpackage", | ||
| tokenizer=Path(args.tokenizer) if args.tokenizer else base / "tokenizer.json", |
There was a problem hiding this comment.
[P1] Fix streaming CLI default bundle and tensor names
The streaming helper assumes the CoreML packages use the legacy names Melspectrogram.mlpackage, ParakeetEncoder.mlpackage, … and that the preprocessor/encoder I/O tensors are called melspectrogram and encoder_output (see _resolve_bundle and _preprocess). However convert-parakeet.py in the same directory now writes files named parakeet_preprocessor.mlpackage, parakeet_encoder.mlpackage, etc. with outputs named mel and encoder, so running speech_to_text_streaming_infer_rnnt.py --model-dir parakeet_coreml will fail immediately with FileNotFoundError or KeyError. The default paths and tensor names here should match the converter output to make the streaming script usable without manually renaming files or writing custom wrappers.
Useful? React with 👍 / 👎.
* docs: add ANE float16 beeping issue to iOS CoreML issues Add issue #7 documenting the Mimi decoder beeping artifact caused by ANE float16 precision loss in streaming state feedback. Update issue #4 to reflect the per-model compute unit configuration. * docs: add voice duration differences issue, update compute unit docs - Add issue #8: voice-dependent duration differences between MLX and CoreML (azelma -2.00s, javert +1.92s — model-level behavior, not fixable in Swift) - Update issue #4: all models now use .cpuAndGPU (not per-model config) - Update issue #7: reflect all-models .cpuAndGPU fix
1 participant
pretty much the same as v3 :)