jamiepine/voicebox · error · HTTPException

Model {model_size} is not downloaded yet. Use /generate to t

Error message

Model {model_size} is not downloaded yet. Use /generate to trigger a download.

What it means

The detail string of the same HTTPException as error 35 — the f-string `f"Model {model_size} is not downloaded yet..."` passed to detail= for qwen/qwen_custom_voice/tada engines when _is_model_cached(model_size) is False. Same root cause and recovery; the {model_size} placeholder interpolates the requested size so the client knows which size is missing.

Source

Thrown at backend/backends/__init__.py:554

        await backend.load_model()


async def ensure_model_cached_or_raise(engine: str, model_size: str = "default") -> None:
    """Check if a model is cached, raise HTTPException if not. Used by streaming endpoint."""
    from fastapi import HTTPException

    backend = get_tts_backend_for_engine(engine)
    cfg = None
    for c in get_tts_model_configs():
        if c.engine == engine and c.model_size == model_size:
            cfg = c
            break

    if engine in ("qwen", "qwen_custom_voice", "tada"):
        if not backend._is_model_cached(model_size):
            raise HTTPException(
                status_code=400,
                detail=f"Model {model_size} is not downloaded yet. Use /generate to trigger a download.",
            )
    else:
        if not backend._is_model_cached():
            display = cfg.display_name if cfg else engine
            raise HTTPException(
                status_code=400,
                detail=f"{display} model is not downloaded yet. Use /generate to trigger a download.",
            )


def unload_model_by_config(config: ModelConfig) -> bool:
    """Unload a model given its config. Returns True if it was loaded, False otherwise."""
    from . import get_tts_backend_for_engine
    from ..services import tts, transcribe, llm as llm_service

    if config.engine == "whisper":
        whisper_model = transcribe.get_whisper_model()
        if whisper_model.is_loaded() and whisper_model.model_size == config.model_size:

View on GitHub (pinned to 51f49dea19)

Solutions

Trigger the download via the /generate endpoint for the exact engine+model_size first.
Pre-fetch all needed sizes during onboarding.
Confirm the cache path and that _is_model_cached(model_size) detects the expected file layout.
List cached sizes in the response so clients can choose an available one.

Example fix

# before
detail=f"Model {model_size} is not downloaded yet. Use /generate to trigger a download."
# after
detail=f"Model '{model_size}' for engine '{engine}' is not downloaded yet. Use /generate to trigger a download."

Defensive patterns

Strategy: validation

Validate before calling

def is_model_cached(engine: str, model_size: str) -> bool:
    backend = get_tts_backend_for_engine(engine)
    if engine in ('qwen', 'qwen_custom_voice', 'tada'):
        return backend._is_model_cached(model_size)
    return backend._is_model_cached()

# Before streaming:
if not is_model_cached(engine, model_size):
    await load_engine_model(engine, model_size)

Try / catch

from fastapi import HTTPException
try:
    await ensure_model_cached_or_raise(engine, model_size)
except HTTPException as e:
    if e.status_code == 400:
        await load_engine_model(engine, model_size)
    else:
        raise

Prevention

Check _is_model_cached(model_size) before the streaming call and trigger /generate when missing.
Keep model sizes consistent between the client request and the cache layout.
Pre-fetch non-default sizes the client may request.

When it happens

Trigger: Streaming endpoint invoked for engine in {qwen, qwen_custom_voice, tada} with a model_size whose weights are absent from the cache directory; _is_model_cached(model_size) returns False.

Common situations: Client requests a size (e.g. '0.6B') before downloading it. Cache wiped. Multiple-size engine where only the default size was pre-fetched.

Related errors

AI-assisted analysis of jamiepine/voicebox@51f49dea19 (2026-08-12). Data as JSON: /api/errors/d9570b261c180f4c. Report an issue: GitHub.