invoke-ai/InvokeAI · error · ValueError

Only Qwen3Encoder_Qwen3Encoder_Config models are supported h

Error message

Only Qwen3Encoder_Qwen3Encoder_Config models are supported here.

What it means

This loader handles the Qwen3 text encoder used by Z-Image and requires the config to be Qwen3Encoder_Qwen3Encoder_Config. Any other config type raises this ValueError immediately, since only that config class carries the layout expectations (text_encoder/ and tokenizer/ subfolders, or a standalone text_encoder root).

Source

Thrown at invokeai/backend/model_manager/load/model_loaders/z_image.py:759

        # so no BFL→diffusers conversion is needed here. The transformer has no tied/shared weights,
        # so we expect a complete state dict — any missing key would leave a required parameter on a
        # meta tensor and fail later during device movement or inference. Fail fast here instead.
        missing, unexpected = model.load_state_dict(sd, assign=True, strict=False)
        raise_on_incomplete_sdnq_load("SDNQ Z-Image transformer", missing, unexpected)
        return model


@ModelLoaderRegistry.register(base=BaseModelType.Any, type=ModelType.Qwen3Encoder, format=ModelFormat.Qwen3Encoder)
class Qwen3EncoderLoader(ModelLoader):
    """Class to load standalone Qwen3 Encoder models for Z-Image (directory format)."""

    def _load_model(
        self,
        config: AnyModelConfig,
        submodel_type: Optional[SubModelType] = None,
    ) -> AnyModel:
        if not isinstance(config, Qwen3Encoder_Qwen3Encoder_Config):
            raise ValueError("Only Qwen3Encoder_Qwen3Encoder_Config models are supported here.")

        model_path = Path(config.path)

        # Support both structures:
        # 1. Full model: model_root/text_encoder/ and model_root/tokenizer/
        # 2. Standalone download: model_root/ contains text_encoder files directly
        text_encoder_path = model_path / "text_encoder"
        tokenizer_path = model_path / "tokenizer"

        # Check if this is a standalone text_encoder download (no nested text_encoder folder)
        is_standalone = not text_encoder_path.exists() and (model_path / "config.json").exists()

        if is_standalone:
            text_encoder_path = model_path
            tokenizer_path = model_path  # Tokenizer files should also be in root

        match submodel_type:
            case SubModelType.Tokenizer:

View on GitHub (pinned to 0b6a024f2f)

Solutions

1. Register the encoder-source model so it resolves to Qwen3Encoder_Qwen3Encoder_Config (standalone Qwen3 encoder download or full model with text_encoder/ + tokenizer/).
2. Update the Z-Image pipeline settings to reference the correctly-typed Qwen3 encoder model.
3. Re-scan the model directory so the model manager classifies the model with the right config class.
4. In custom code, gate the call with isinstance(config, Qwen3Encoder_Qwen3Encoder_Config).

Example fix

// before
config = Main_Checkpoint_ZImage_Config(path=p)
enc = qwen3_loader._load_model(config, SubModelType.TextEncoder)  # ValueError
// after
config = Qwen3Encoder_Qwen3Encoder_Config(path=p)  # p is the Qwen3 encoder root
enc = qwen3_loader._load_model(config, SubModelType.TextEncoder)

if not isinstance(config, Qwen3Encoder_Qwen3Encoder_Config):
    raise ValueError(f"Qwen3 encoder loader requires Qwen3Encoder_Qwen3Encoder_Config, got {type(config).__name__}")

def is_qwen3_encoder_config(config: AnyModelConfig) -> bool:
    return isinstance(config, Qwen3Encoder_Qwen3Encoder_Config)

try:
    enc = qwen3_loader._load_model(config, SubModelType.TextEncoder)
except ValueError as e:
    if "Qwen3Encoder_Qwen3Encoder_Config" in str(e):
        config = reclassify_as_qwen3_encoder(config.path)
        enc = qwen3_loader._load_model(config, SubModelType.TextEncoder)
    else:
        raise

When it happens

Trigger: Calling _load_model with a config that is not Qwen3Encoder_Qwen3Encoder_Config — e.g. pointing the VAE/text-encoder source model at a checkpoint, GGUF, or SDNQ config — fails the isinstance check at z_image.py:759.

Common situations: The 'Qwen3 & VAE source model' referenced by a Z-Image pipeline was registered with the wrong model type; a full Z-Image checkpoint was selected as the encoder source instead of a Qwen3 encoder model; duplicate/mis-typed model records after re-import.

Related errors

• Expected Main_Checkpoint_ZImage_Config, got {type(config).__
• Expected Main_GGUF_ZImage_Config, got {type(config).__name__
• Expected PreTrainedModel for text encoder, got {type(text_en
• Expected PreTrainedTokenizerBase for tokenizer, got {type(to
• Only MistralEncoder_Diffusers_Config models are supported he

AI-assisted analysis of invoke-ai/InvokeAI@0b6a024f2f (2026-08-29). Data as JSON: /api/errors/8dc71c3d2e480403. Report an issue: GitHub.