mx.concat/mx.concatenate incompatible with MLX 0.30.6 (axis keyword arg rejected)

[JarodTan01]

Description

With parakeet-mlx==0.5.0 and mlx==0.30.6, streaming transcription via transcribe_stream() / add_audio() fails immediately with a TypeError from MLX's concat/concatenate functions.

Error

TypeError: concat(): incompatible function arguments. The following argument types are supported:
    1. concat(arrays: list[array], axis: Optional[int] = 0, *, stream: Union[None, Stream, Device] = None) -> array

Invoked with types: list, kwargs = { axis: int }

Root Cause

Two issues:

1. mx.concat is removed/broken in MLX 0.30.6 — should use mx.concatenate instead. Affected files:

   • parakeet.py:1004 — mx.concat([self.audio_buffer, audio], axis=0)
   • parakeet.py:1025 — mx.concat([self.mel_buffer, mel], axis=1)
   • audio.py:146 — mx.concat([x[:1], ...], axis=0)
   • conformer.py:293 — mx.concat([...])

2. axis cannot be passed as a keyword argument in MLX 0.30.6's pybind11 bindings — it must be positional. All calls like mx.concatenate([...], axis=N) need to become mx.concatenate([...], N). Affected files:

   • cache.py — multiple occurrences with axis=1 and axis=2

Additionally, add_audio() in parakeet.py receives numpy arrays from audio capture, but mx.concatenate can't handle mixed np.ndarray / mx.array lists. Adding if not isinstance(audio, mx.array): audio = mx.array(audio) before the concatenate call fixes this.

Environment

• macOS 15 (Apple Silicon)
• Python 3.14.0
• mlx==0.30.6
• parakeet-mlx==0.5.0

Workaround

Replace all mx.concat → mx.concatenate, change axis=N keyword args to positional N, and convert numpy input to mx.array in add_audio().

[senstella]

Hello, thank you for creating the issue. Unfortunately, I couldn't replicate the issues you've mentioned. Would you mind providing a reproducible script? I'm pretty sure concat is not removed! See the main branch of MLX.

Additionally, add_audio() in parakeet.py receives numpy arrays from audio capture, but mx.concatenate can't handle mixed np.ndarray / mx.array lists. Adding if not isinstance(audio, mx.array): audio = mx.array(audio) before the concatenate call fixes this.

Also, add_audio in parakeet.py have mx.array as an input already.

    def add_audio(self, audio: mx.array) -> None:
        """Takes portion of audio and transcribe it.

        `audio` must be 1D array"""

If you're passing a numpy array on add_audio(), you should convert it to MLX array via mx.array()!