unslothai/unsloth · warning · RuntimeError

load error: {p.get('error')}

Error message

load error: {p.get('error')}

What it means

Thrown by AudioAttachmentAdapter.add (audio-attachment-adapter.ts:47-60) when a model IS loaded and not loading, but the active model lacks hasAudioInput. The message names the offending model via its display name, externalModelLabel(checkpoint) for external:: ids (issue #8405), or the raw checkpoint string. It fires at file-pick time, mirroring the image adapter gate, and is accompanied by a toast.

Source

Thrown at scripts/diffusion_bench.py:172

# ── load + generate ────────────────────────────────────────────────────────


def _wait_for_load(backend: Any, timeout_s: int = 2400) -> None:
    deadline = time.time() + timeout_s
    last = None
    while time.time() < deadline:
        p = backend.load_progress()
        phase = p.get("phase")
        if phase != last:
            last = phase
            frac = p.get("fraction") or 0.0
            bt = (p.get("bytes_total") or 0) / 1e9
            print(f"  load phase={phase} frac={frac:.3f} total={bt:.2f}GB", flush = True)
        if phase == "ready":
            return
        if phase == "error":
            raise RuntimeError(f"load error: {p.get('error')}")
        time.sleep(2)
    raise TimeoutError(f"model load did not reach ready within {timeout_s}s")


def _generate_once(backend: Any, args: argparse.Namespace) -> Any:
    """One generation at the fixed seed; returns the first PIL image."""
    result = backend.generate(
        prompt = args.prompt,
        width = args.width,
        height = args.height,
        steps = args.steps,
        guidance = args.guidance,
        seed = args.seed,
        batch_size = args.batch_size,
    )
    images = result["images"]
    return images[0]

View on GitHub (pinned to 203007d190)

Solutions

1. Load a model that accepts audio input (e.g. an audio-capable multimodal model) before attaching audio files.
2. If using an external provider, pick a model known to accept audio (e.g. a provider model with an audio tier); verify its capability metadata in the models list.
3. Remove the audio attachment and send the message as text-only if audio is not essential.
4. As a developer adding provider models, ensure the model registry entry sets hasAudioInput so the gate can recognize it.

function modelAcceptsAudio(state: { params: { checkpoint: string }; models: Array<{id: string; hasAudioInput?: boolean}>; modelLoading: boolean }): boolean {
  const m = state.models.find((x) => x.id === state.params.checkpoint);
  return !!state.params.checkpoint && !state.modelLoading && !!m?.hasAudioInput;
}

try {
  await adapter.add({ file });
} catch (error) {
  // adapter already toasts; swallow to avoid a double toast
  if (!(error instanceof Error && /cannot accept audio/.test(error.message))) throw error;
}

When it happens

Trigger: User opens the 'Add photos & files' picker and selects an audio file while the loaded checkpoint is a text/vision-only model or an external provider model whose registry entry lacks audio input capability. The check is activeModel?.hasAudioInput === undefined/false — including external models that have no row in state.models at all.

Common situations: Loading a small local LLM (no audio tower) and trying to attach a voice memo; switching to an OpenAI-compatible provider whose model metadata does not advertise audio; using a Whisper-typed LoRA where the base checkpoint id resolves to a non-audio row.

Related errors

• model load did not reach ready within {timeout_s}s
• load error: {p.get('error')}
• model load did not reach ready
• unknown family '{name}'
• Refused notebook fetch from {host!r}: not in allowlist {sort

AI-assisted analysis of unslothai/unsloth@203007d190 (2026-08-15). Data as JSON: /api/errors/6c5cad7bdb2ea0a4. Report an issue: GitHub.