invoke-ai/InvokeAI · error · TypeError

Expected PreTrainedModel for text encoder, got {type(text_en

Error message

Expected PreTrainedModel for text encoder, got {type(text_encoder).__name__}. The Qwen3 encoder model may be corrupted or incompatible.

What it means

InvokeAI's FLUX.2 Klein text encoder invocation requires the loaded Qwen3 text encoder to be a transformers PreTrainedModel instance. If the object loaded from the model manager is any other type, the model file is likely corrupted, partially downloaded, or is not actually a Qwen3 encoder (incompatible model directory). The check guards downstream forward passes that assume the transformers API.

Source

Thrown at invokeai/app/invocations/flux2_klein_text_encoder.py:136

                f"Recovered {repaired_tensors} required Qwen3 tensor(s) onto {device} after a partial device mismatch."
            )

        # Apply LoRA models
        lora_dtype = TorchDevice.choose_bfloat16_safe_dtype(device)
        exit_stack.enter_context(
            LayerPatcher.apply_smart_model_patches(
                model=text_encoder,
                patches=self._lora_iterator(context),
                prefix=FLUX_LORA_T5_PREFIX,
                dtype=lora_dtype,
                cached_weights=cached_weights,
            )
        )

        context.util.signal_progress("Running Qwen3 text encoder (Klein)")

        if not isinstance(text_encoder, PreTrainedModel):
            raise TypeError(
                f"Expected PreTrainedModel for text encoder, got {type(text_encoder).__name__}. "
                "The Qwen3 encoder model may be corrupted or incompatible."
            )
        if not isinstance(tokenizer, PreTrainedTokenizerBase):
            raise TypeError(
                f"Expected PreTrainedTokenizerBase for tokenizer, got {type(tokenizer).__name__}. "
                "The Qwen3 tokenizer may be corrupted or incompatible."
            )

        messages = [{"role": "user", "content": prompt}]

        text: str = tokenizer.apply_chat_template(  # type: ignore[assignment]
            messages,
            tokenize=False,
            add_generation_prompt=True,
            enable_thinking=False,
        )

View on GitHub (pinned to 0b6a024f2f)

Solutions

Re-download the Qwen3 text encoder model (delete the model folder and re-add it in InvokeAI model manager)
Verify the model config points to a genuine Qwen3 encoder directory containing config.json and safetensors weights
Check that installed transformers version supports Qwen3 and returns PreTrainedModel instances
Re-run 'invokeai-migrate' / model scan to repair broken model records

Example fix

// before (corrupt/mismatched encoder)
ModelConfig(type='main', base='Flux2', path='/models/qwen3-encoder-wrong/')
// after
ModelConfig(type='main', base='Flux2', path='/models/Qwen/Qwen3-encoder/', name='Qwen3 encoder')

Defensive patterns

Strategy: type-guard

Validate before calling

from invokeai.backend.model_manager.load import ModelLoaderRegistry
info = context.models.load(qwen3_encoder.text_encoder)
if not isinstance(info.model, PreTrainedModel):
    raise TypeError(f'Qwen3 encoder invalid: {type(info.model).__name__}')

Type guard

from transformers import PreTrainedModel

def is_qwen3_encoder(obj) -> bool:
    return isinstance(obj, PreTrainedModel)

Try / catch

try:
    result = klein_encoder.invoke(context)
except TypeError as e:
    if 'PreTrainedModel for text encoder' in str(e):
        reimport_model_manager_entry(qwen3_encoder.text_encoder)
    raise

Prevention

Verify model integrity hashes after download
Keep transformers updated to a Qwen3-compatible version
Point configs only at verified Qwen3 encoder directories
Re-import models that fail integrity checks

When it happens

Trigger: context.models.load() returns an object whose class is not PreTrainedModel when _encode_prompt places the Qwen3 encoder on device; e.g. the model folder points to a non-Qwen model, files are truncated/corrupt, or a loader fallback returned a raw module.

Common situations: Interrupted model downloads leaving incomplete safetensors files; users pointing a FLUX.2 Klein model config at the wrong encoder directory; transformers version changes causing a different wrapper class to be loaded; hash-mismatched model installs.

Related errors

AI-assisted analysis of invoke-ai/InvokeAI@0b6a024f2f (2026-08-29). Data as JSON: /api/errors/c52c90fd69d9dccc. Report an issue: GitHub.