invoke-ai/InvokeAI · error · NotAMatchError

model architecture '{class_name}' is not a causal language m

Error message

model architecture '{class_name}' is not a causal language model

What it means

NotAMatchError from TextLLM from_model_on_disk. The generic text-LLM config only accepts causal language model architectures: the config dict's architecture class name must end with 'ForCausalLM' (LlamaForCausalLM, Phi3ForCausalLM, Qwen2ForCausalLM, etc.). Anything else (encoders, seq2seq, vision models) is rejected as not a supported causal LM.

Source

Thrown at invokeai/backend/model_manager/configs/text_llm.py:48

    """Model config for text-only causal language models (e.g. Llama, Phi, Qwen, Mistral)."""

    type: Literal[ModelType.TextLLM] = Field(default=ModelType.TextLLM)
    base: Literal[BaseModelType.Any] = Field(default=BaseModelType.Any)
    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:
        raise_if_not_dir(mod)

        raise_for_override_fields(cls, override_fields)

        # Check that the model's architecture is a causal language model.
        # This covers LlamaForCausalLM, PhiForCausalLM, Phi3ForCausalLM, Qwen2ForCausalLM,
        # MistralForCausalLM, GemmaForCausalLM, GPTNeoXForCausalLM, etc.
        config_dict = get_config_dict_or_raise(common_config_paths(mod.path))
        class_name = get_class_name_from_config_dict_or_raise(config_dict)
        if not class_name.endswith("ForCausalLM"):
            raise NotAMatchError(f"model architecture '{class_name}' is not a causal language model")

        # During *automatic* classification, defer to the dedicated PiD Gemma2 encoder config — but only
        # for the hidden size that config actually accepts (2304 = Gemma-2-2b). Larger Gemma 2 variants
        # (9B=3584, 27B=4608) are rejected by the encoder config, so they must remain classifiable as a
        # generic TextLLM here rather than falling through to Unknown. An explicit `type=text_llm` request
        # always keeps the model as TextLLM (the generic AutoModelForCausalLM loader supports these).
        explicitly_requested_text_llm = override_fields.get("type") == ModelType.TextLLM
        if (
            not explicitly_requested_text_llm
            and class_name == "Gemma2ForCausalLM"
            and config_dict.get("hidden_size") == _GEMMA2_2B_HIDDEN_SIZE
        ):
            raise NotAMatchError(
                "architecture 'Gemma2ForCausalLM' (2304-dim Gemma-2-2b) is handled by the PiD encoder config, not TextLLM"
            )

        # Verify tokenizer files exist to avoid runtime failures
        tokenizer_files = {"tokenizer.json", "tokenizer.model", "tokenizer_config.json"}

View on GitHub (pinned to 0b6a024f2f)

Solutions

Use a model with a *ForCausalLM architecture (Llama, Qwen2, Phi3, Mistral, Gemma, GPTNeoX families)
Fix config.json 'architectures' if it was corrupted, using the value from the upstream repo
Register encoder/seq2seq models under their appropriate InvokeAI model type instead of text_llm

Example fix

// before
{"architectures": ["T5EncoderModel"], ...}
// after
{"architectures": ["Qwen2ForCausalLM"], ...}  # or use a causal-LM model

Defensive patterns

Strategy: validation

Validate before calling

import json
def is_causal_lm(model_dir) -> bool:
    cfg = json.loads((model_dir / "config.json").read_text())
    archs = cfg.get("architectures", [])
    return bool(archs) and archs[0].endswith("ForCausalLM")

Try / catch

try:
    install_model(path, type="text_llm")
except NotAMatchError as e:
    if "causal language model" in str(e):
        logger.error("Architecture is not *ForCausalLM; choose an encoder/seq2seq-appropriate type")

Prevention

Check config.json 'architectures' before adding a model as text_llm
Only use *ForCausalLM models (Llama/Qwen2/Phi3/Mistral/Gemma) as text LLMs
Don't hand-edit config.json architectures

When it happens

Trigger: from_model_on_disk on a model whose config.json 'architectures' class does not end with 'ForCausalLM' — e.g. T5ForConditionalGeneration, BertModel, CLIPTextModel, or a malformed config.json missing 'architectures'.

Common situations: Trying to load encoder-only or encoder-decoder models (T5, BERT, CLIP) as chat LLMs, embedding models, or configs whose architectures entry was edited/stripped.

Related errors

AI-assisted analysis of invoke-ai/InvokeAI@0b6a024f2f (2026-08-29). Data as JSON: /api/errors/211ad0fb95a38755. Report an issue: GitHub.