unslothai/unsloth · warning · AudioBackendUnsupportedError

audio_unsupported_backend

audio_unsupported_backend

Error message

This backend cannot generate audio.

What it means

AudioBackendUnsupportedError with code 'audio_unsupported_backend': the active model's inference backend has no code path for this TTS task. The source comment is explicit — a tagged code means 'no path for this task, not a failure'; the error may carry a worker-provided hint about what backend would work.

Source

Thrown at studio/backend/core/inference/orchestrator.py:2226

                            self._mark_worker_started(cancel_event)
                            worker_started = True
                            continue

                        if rtype == "audio_done":
                            if cancel_event is not None and cancel_event.is_set():
                                raise AudioGenerationCancelledError("Audio generation cancelled")
                            wav_bytes = base64.b64decode(resp["wav_base64"])
                            sample_rate = resp["sample_rate"]
                            return wav_bytes, sample_rate

                        if rtype == "audio_error":
                            if resp.get("cancelled") or (
                                cancel_event is not None and cancel_event.is_set()
                            ):
                                raise AudioGenerationCancelledError("Audio generation cancelled")
                            # Tagged code = no path for this task, not a failure.
                            if resp.get("code") == AUDIO_UNSUPPORTED_CODE:
                                raise AudioBackendUnsupportedError(
                                    resp.get("error", "This backend cannot generate audio."),
                                    hint = resp.get("hint"),
                                )
                            raise RuntimeError(resp.get("error", "Audio generation failed"))

                        if rtype == "error":
                            if cancel_event is not None and cancel_event.is_set():
                                raise AudioGenerationCancelledError("Audio generation cancelled")
                            raise RuntimeError(resp.get("error", "Unknown error"))

                        if rtype == "status":
                            continue

                    # A caller cancellation already spent the drain window polling this
                    # request's mailbox. Tear down an unresponsive worker now instead of
                    # waiting out the much longer generation watchdog or draining twice.
                    if cancel_deadline is not None:
                        if self._shutdown_subprocess(timeout = _AUDIO_CANCEL_DRAIN_TIMEOUT):

View on GitHub (pinned to 203007d190)

Solutions

1. Load a TTS-capable model/backend before requesting audio.
2. Read the hint field — it names the supported backend or model shape.
3. Disable the audio action in the UI for non-TTS models (use this error's code to detect the condition).

Example fix

// before
wav, sr = orchestrator.generate_audio(text)
// after
from core.inference.orchestrator import AudioBackendUnsupportedError
try:
    wav, sr = orchestrator.generate_audio(text)
except AudioBackendUnsupportedError as e:
    show_info(e.hint or "This model cannot generate audio")

if not model_supports_tts(orchestrator.active_model_name):
    disable_audio_button()

from core.inference.orchestrator import AudioBackendUnsupportedError

def is_audio_unsupported(exc: Exception) -> bool:
    return isinstance(exc, AudioBackendUnsupportedError) and getattr(exc, "code", None) == "audio_unsupported_backend"

try:
    wav, sr = orchestrator.generate_audio(text)
except AudioBackendUnsupportedError as e:
    show_capability_message(e.hint)

When it happens

Trigger: Calling generate_audio on a model loaded with a text-only backend (e.g. a plain transformers LLM without TTS support), or requesting an audio modality the loaded adapter/backend cannot produce.

Common situations: UI allows the audio button for models without TTS support; user loads a chat model and triggers speech generation; backend matrix mismatch after a model switch.

Related errors

• Inference subprocess is not running
• No active model
• Audio generation cancelled
• No valid audio codes found after START_OF_SPEECH token
• No bicodec_semantic tokens found in generated output

AI-assisted analysis of unslothai/unsloth@203007d190 (2026-08-15). Data as JSON: /api/errors/de7b2f6eb450f289. Report an issue: GitHub.