invoke-ai/InvokeAI · error · TypeError

Expected Qwen3Encoder_Checkpoint_Config, got {type(config)._

Error message

Expected Qwen3Encoder_Checkpoint_Config, got {type(config).__name__}. Model configuration type mismatch.

What it means

_load_from_singlefile re-validates the config with a strict isinstance check against Qwen3Encoder_Checkpoint_Config and raises TypeError naming the actual received type if it fails. This is a defense-in-depth check: the dispatching _load_model already guards, so hitting this means the loader was invoked through an unusual path or the config type changed between checks.

Source

Thrown at invokeai/backend/model_manager/load/model_loaders/z_image.py:948

        tokenizer is identical across the 0.6B / 4B / 8B variants, so we load the
        self-contained copy vendored in the package — fully offline, no HuggingFace
        download required.
        """
        return load_bundled_qwen3_tokenizer()

    def _load_from_singlefile(
        self,
        config: AnyModelConfig,
    ) -> AnyModel:
        from safetensors.torch import load_file
        from transformers import Qwen3Config, Qwen3ForCausalLM

        from invokeai.backend.util.logging import InvokeAILogger

        logger = InvokeAILogger.get_logger(self.__class__.__name__)

        if not isinstance(config, Qwen3Encoder_Checkpoint_Config):
            raise TypeError(
                f"Expected Qwen3Encoder_Checkpoint_Config, got {type(config).__name__}. "
                "Model configuration type mismatch."
            )
        model_path = Path(config.path)

        # Determine safe dtype based on target device capabilities
        target_device = TorchDevice.choose_torch_device()
        model_dtype = TorchDevice.choose_bfloat16_safe_dtype(target_device)

        # Load the state dict from safetensors file
        sd = load_file(model_path)

        # Handle ComfyUI quantized checkpoints
        # ComfyUI stores quantized weights with accompanying scale factors:
        # - layer.weight: quantized data (FP8)
        # - layer.weight_scale: scale factor (FP32 scalar)
        # Dequantization formula: dequantized = weight.to(dtype) * weight_scale
        # Reference: https://github.com/Comfy-Org/ComfyUI/blob/master/QUANTIZATION.md

View on GitHub (pinned to 0b6a024f2f)

Solutions

Pass a genuine Qwen3Encoder_Checkpoint_Config instance (read the printed type name in the message to see what you actually passed).
Re-create the model record via the model manager so the correct config class is instantiated.
If calling internals directly, build the config via the same factory the registry uses.
Upgrade/downgrade consistently — do not mix config definitions from different InvokeAI versions.

Example fix

// before
self._load_from_singlefile(config_dict)
// after
from invokeai.backend.model_manager.load.model_loaders.z_image import Qwen3Encoder_Checkpoint_Config
config = Qwen3Encoder_Checkpoint_Config(**config_dict)
self._load_from_singlefile(config)

Defensive patterns

Strategy: type-guard

Validate before calling

if not isinstance(config, Qwen3Encoder_Checkpoint_Config):
    raise TypeError(f"Cannot single-file load with {type(config).__name__}")

Type guard

def is_qwen3_ckpt(c: AnyModelConfig) -> bool:
    return isinstance(c, Qwen3Encoder_Checkpoint_Config)

Try / catch

try:
    model = loader._load_from_singlefile(config)
except TypeError as e:
    if "Expected Qwen3Encoder_Checkpoint_Config" in str(e):
        config = Qwen3Encoder_Checkpoint_Config(**vars(config))
        model = loader._load_from_singlefile(config)
    else:
        raise

Prevention

Use the model manager/registry to build configs, not raw dicts.
Call _load_model rather than _load_from_singlefile directly.
Don't persist and reuse config blobs across InvokeAI version upgrades without re-instantiation.
Read the received type name in the message to diagnose quickly.

When it happens

Trigger: Calling _load_from_singlefile directly with a non-Qwen3Encoder_Checkpoint_Config object; a custom subclass or wrapper passes a duck-typed config that fails the exact isinstance check; internal refactors bypass the _load_model guard.

Common situations: Scripting the loader internals instead of using the model manager; config objects deserialized from cache/DB losing their concrete class; mixing config classes across InvokeAI versions after an upgrade.

Related errors

AI-assisted analysis of invoke-ai/InvokeAI@0b6a024f2f (2026-08-29). Data as JSON: /api/errors/870934b90ec2c81d. Report an issue: GitHub.