invoke-ai/InvokeAI · error · NotAMatchError

directory does not contain Gemma2 tokenizer files (tokenizer

Error message

directory does not contain Gemma2 tokenizer files (tokenizer.json/tokenizer.model)

What it means

NotAMatchError raised by Gemma2Encoder_Gemma2Encoder_Config.from_model_on_disk when the model directory contains a valid Gemma2ForCausalLM config.json but no tokenizer.json or tokenizer.model file. PiD later calls AutoTokenizer.from_pretrained on this same directory, so the tokenizer files must live alongside the weights; the config refuses to match a directory that would fail at load time.

Source

Thrown at invokeai/backend/model_manager/configs/gemma2_encoder.py:85

            raise NotAMatchError("directory looks like a full diffusers pipeline, not a standalone Gemma2 encoder")

        # Architecture marker is the canonical signal.
        raise_for_class_name(config_path, {"Gemma2ForCausalLM"})

        # Only Gemma-2-2b (2304-dim hidden state) is compatible with PiD's fixed caption projection.
        # Reject 9B/27B variants here so they are not offered as compatible encoders and then fail with
        # a matrix-shape error deep inside PiD inference.
        hidden_size = get_config_dict_or_raise(config_path).get("hidden_size")
        if hidden_size != _PID_GEMMA_HIDDEN_SIZE:
            raise NotAMatchError(
                f"Gemma2 hidden_size {hidden_size} is incompatible with PiD, which requires "
                f"{_PID_GEMMA_HIDDEN_SIZE} (Gemma-2-2b); 9B/27B variants are not supported."
            )

        # Sanity check that tokenizer files live alongside the model (PiD calls
        # AutoTokenizer.from_pretrained on the same directory).
        if not any((mod.path / f).exists() for f in ("tokenizer.json", "tokenizer.model")):
            raise NotAMatchError("directory does not contain Gemma2 tokenizer files (tokenizer.json/tokenizer.model)")

        return cls(**override_fields)


def _read_gguf_arch_and_hidden_size(path: Path) -> tuple[str, int | None]:
    """Read (general.architecture, <arch>.embedding_length) from a GGUF file's metadata.

    Raises NotAMatchError if the file is not a readable GGUF or is missing the architecture marker.
    """
    import gguf

    try:
        reader = gguf.GGUFReader(path)
    except Exception as e:
        raise NotAMatchError(f"not a readable GGUF file: {e}") from e

    arch_field = reader.fields.get("general.architecture")
    if arch_field is None:

View on GitHub (pinned to 0b6a024f2f)

Solutions

Copy tokenizer.json (or tokenizer.model) and tokenizer_config.json from the original HuggingFace repo into the model directory root
Re-download the full gemma-2-2b-it repo rather than cherry-picking files
Verify with ls: the directory must contain config.json, model*.safetensors, and tokenizer.json/tokenizer.model at the same level

Example fix

// before
models/gemma-2-2b-it/{config.json, model.safetensors}
// after
models/gemma-2-2b-it/{config.json, model.safetensors, tokenizer.json, tokenizer_config.json}

Defensive patterns

Strategy: validation

Validate before calling

from pathlib import Path

def tokenizer_files_present(model_dir: str | Path) -> bool:
    d = Path(model_dir)
    return any((d / f).exists() for f in ("tokenizer.json", "tokenizer.model"))

Type guard

def is_complete_gemma2_dir(p: Path) -> bool:
    return p.is_dir() and (p / "config.json").is_file() and tokenizer_files_present(p)

Try / catch

try:
    add_model_and_import(model_dir)
except NotAMatchError as e:
    if "tokenizer" in str(e):
        print("Copy tokenizer.json/tokenizer.model from the HuggingFace repo into the model dir")

Prevention

Download the whole HuggingFace repo (huggingface-cli download) instead of cherry-picking files
After copying models, verify config.json + weights + tokenizer.json/tokenizer.model sit at the same directory level
Keep tokenizer_config.json too — PiD calls AutoTokenizer.from_pretrained on the directory

When it happens

Trigger: Importing a directory that has config.json + weights but the tokenizer files were deleted, not downloaded, or were saved into a sibling subdirectory instead of the model root.

Common situations: Manually copying only the safetensors and config.json from a HuggingFace repo; a download tool that skips tokenizer assets; hand-assembled model folders; using save_pretrained on a subdirectory while pointing InvokeAI at the parent.

Related errors

AI-assisted analysis of invoke-ai/InvokeAI@0b6a024f2f (2026-08-29). Data as JSON: /api/errors/7c3d5b8a0b611a5d. Report an issue: GitHub.