invoke-ai/InvokeAI · error · NotAMatchError

Krea-2 requires a Qwen3-VL 4B checkpoint containing language

Error message

Krea-2 requires a Qwen3-VL 4B checkpoint containing language-model layer 35

What it means

After checking hidden size, the same validator requires the state dict to contain a language-model layer with index 35 (".layers.35."), which only exists in the full 36-layer Qwen3-VL 4B decoder. This NotAMatchError is thrown when layer 35 is missing, indicating the checkpoint is a smaller/trimmed Qwen3-VL model or a truncated weight file rather than the required 4B checkpoint.

Source

Thrown at invokeai/backend/model_manager/configs/qwen3_vl_encoder.py:84


def _validate_krea2_qwen3_vl_checkpoint_shape(state_dict: dict[str | int, Any]) -> None:
    embed_keys = (
        "model.embed_tokens.weight",
        "model.language_model.embed_tokens.weight",
        "language_model.embed_tokens.weight",
        "embed_tokens.weight",
    )
    embed = next((state_dict[key] for key in embed_keys if key in state_dict), None)
    shape = getattr(embed, "shape", ())
    if len(shape) < 2 or shape[1] != _KREA2_QWEN3_VL_HIDDEN_SIZE:
        hidden_size = shape[1] if len(shape) >= 2 else None
        raise NotAMatchError(
            f"Krea-2 requires a Qwen3-VL 4B checkpoint with hidden size "
            f"{_KREA2_QWEN3_VL_HIDDEN_SIZE}, got {hidden_size}"
        )
    if not any(isinstance(key, str) and ".layers.35." in key for key in state_dict):
        raise NotAMatchError("Krea-2 requires a Qwen3-VL 4B checkpoint containing language-model layer 35")


class Qwen3VLEncoder_Qwen3VLEncoder_Config(Config_Base):
    """Configuration for standalone Qwen3-VL text encoder models (diffusers-like directory format).

    Used by Krea-2, whose text conditioning comes from a Qwen3-VL model (``Qwen3VLModel``). The model
    weights are expected either in a ``text_encoder`` subfolder of the model directory or directly at the
    root (standalone download). This is distinct from the text-only ``Qwen3Encoder`` (Z-Image / FLUX.2
    Klein) and the Qwen2.5-VL ``QwenVLEncoder`` (Qwen Image).
    """

    base: Literal[BaseModelType.Any] = Field(default=BaseModelType.Any)
    type: Literal[ModelType.Qwen3VLEncoder] = Field(default=ModelType.Qwen3VLEncoder)
    format: Literal[ModelFormat.Qwen3VLEncoder] = Field(default=ModelFormat.Qwen3VLEncoder)
    cpu_only: bool | None = Field(default=None, description="Whether this model should run on CPU only")

    @classmethod
    def from_model_on_disk(cls, mod: ModelOnDisk, override_fields: dict[str, Any]) -> Self:

View on GitHub (pinned to 0b6a024f2f)

Solutions

Obtain the complete Qwen3-VL 4B checkpoint (36 layers) referenced by Krea-2 and re-import it.
List the state-dict keys (e.g. with safetensors.safe_open) and confirm keys like model.language_model.layers.35.* exist; if not, re-download missing shards.
If the model is sharded, ensure all shards and the index file are present and fully downloaded before scanning the folder.
If you intentionally use a trimmed model, register it under a different model type; InvokeAI will not accept it as the Krea-2 encoder.

Example fix

// before: trimmed checkpoint keys end at layer 27
model.layers.27.self_attn.q_proj.weight
// after: full 4B checkpoint contains layers 0..35
model.layers.35.mlp.down_proj.weight  # required by validator

Defensive patterns

Strategy: validation

Validate before calling

from safetensors import safe_open

def has_all_36_layers(path: str) -> bool:
    with safe_open(path, framework="pt") as f:
        keys = f.keys()
        return any(".layers.35." in k for k in keys)

Type guard

def is_full_depth_qwen3vl(sd: dict) -> bool:
    return any(isinstance(k, str) and ".layers.35." in k for k in sd)

Try / catch

try:
    cfg = Qwen3VLEncoder_Checkpoint_Config.from_model_on_disk(mod, {})
except NotAMatchError:
    logger.warning("Checkpoint %s lacks layer 35; expected the full 36-layer Qwen3-VL 4B model", mod.path)

Prevention

Confirm the model card states 36 layers (num_hidden_layers=36) before downloading a Qwen3-VL encoder for Krea-2.
For sharded downloads, ensure every shard finished downloading before importing.
Avoid pruned/layer-dropped community reuploads of Qwen3-VL.
Inspect state-dict key ranges (layers.0 ... layers.35) before adding a checkpoint to the models folder.

When it happens

Trigger: Qwen3VLEncoder_Checkpoint_Config.from_model_on_disk loads a .safetensors state dict that has an embed token and a visual tower, but no key matching ".layers.35." (0-indexed layer 35 of 36), then calls _validate_krea2_qwen3_vl_checkpoint_shape.

Common situations: Using a smaller Qwen3-VL variant (fewer layers, e.g. 2B with ~28 layers), a pruned or layer-dropped distillation, a sharded download where the shard containing the final layers was not fully downloaded/converted, or hand-built state dicts that omit trailing layers.

Related errors

AI-assisted analysis of invoke-ai/InvokeAI@0b6a024f2f (2026-08-29). Data as JSON: /api/errors/5cacd1ddaca9509d. Report an issue: GitHub.