invoke-ai/InvokeAI · error · TypeError

Expected PreTrainedTokenizerBase for Gemma tokenizer, got {t

Error message

Expected PreTrainedTokenizerBase for Gemma tokenizer, got {type(gemma_tokenizer).__name__}.

What it means

The same Gemma loading path also materializes the tokenizer and requires it to be a HuggingFace PreTrainedTokenizerBase before encoding prompts. A non-tokenizer object means the record referenced by gemma2_encoder.tokenizer is wrong or was loaded incorrectly, so a TypeError is raised before any caption encoding.

Source

Thrown at invokeai/app/invocations/sdxl_pid_decode.py:145

                    scaling_factor = float(getattr(vae, "scale_factor", scaling_factor))
                    shift_factor = float(getattr(vae, "shift_factor", shift_factor))
            del vae_info
            TorchDevice.empty_cache()
        context.logger.info(
            f"SDXL PiD decode: latent shape={tuple(latents.shape)} (expect [B, 4, H/8, W/8]) dtype={latents.dtype} "
            f"using scale={scaling_factor:.5f} shift={shift_factor:.5f}"
        )

        # 2) Encode caption with Gemma-2.
        gemma_text_encoder_info = context.models.load(self.gemma2_encoder.text_encoder)
        gemma_tokenizer_info = context.models.load(self.gemma2_encoder.tokenizer)
        with ExitStack() as stack:
            (_, gemma_encoder) = stack.enter_context(gemma_text_encoder_info.model_on_device())
            (_, gemma_tokenizer) = stack.enter_context(gemma_tokenizer_info.model_on_device())
            if not isinstance(gemma_encoder, PreTrainedModel):
                raise TypeError(f"Expected PreTrainedModel for Gemma encoder, got {type(gemma_encoder).__name__}.")
            if not isinstance(gemma_tokenizer, PreTrainedTokenizerBase):
                raise TypeError(
                    f"Expected PreTrainedTokenizerBase for Gemma tokenizer, got {type(gemma_tokenizer).__name__}."
                )

            # Encode on the encoder's intended compute device. compute_device honours cpu_only and is
            # stable under partial loading — the first parameter may be offloaded to CPU while later
            # modules load on CUDA, so inferring the device from the first parameter could place caption
            # inputs on the wrong device.
            device = gemma_text_encoder_info.compute_device
            encode_dtype = TorchDevice.choose_bfloat16_safe_dtype(device)
            context.util.signal_progress("Encoding caption with Gemma-2")
            caption_embs, caption_mask = encode_caption_for_pid(
                [self.prompt],
                tokenizer=gemma_tokenizer,
                encoder=gemma_encoder,
                device=device,
                dtype=encode_dtype,
            )
            caption_embs = caption_embs.detach().to("cpu")

View on GitHub (pinned to 0b6a024f2f)

Solutions

1. Re-point the Gemma2 encoder node's tokenizer field at the correct tokenizer record and re-run.
2. Re-download/re-import the Gemma encoder so tokenizer files and records are rebuilt.
3. Confirm tokenizer files (tokenizer.json/tokenizer.model) exist in the model folder.
4. Verify the installed transformers version matches what InvokeAI expects.

Example fix

// before
(_, gemma_tokenizer) = stack.enter_context(gemma_tokenizer_info.model_on_device())
// after
(_, gemma_tokenizer) = stack.enter_context(gemma_tokenizer_info.model_on_device())
assert isinstance(gemma_tokenizer, PreTrainedTokenizerBase), f"bad tokenizer: {type(gemma_tokenizer).__name__}"

tok_path = gemma_model_dir / "tokenizer.json"
if not tok_path.exists():
    raise FileNotFoundError("Gemma tokenizer files missing")

def is_tokenizer(obj) -> bool:
    from transformers import PreTrainedTokenizerBase
    return isinstance(obj, PreTrainedTokenizerBase)

try:
    result = invoke(context)
except TypeError as e:
    if "Gemma tokenizer" in str(e):
        reimport_gemma_tokenizer()
        retry(context)
    else:
        raise

When it happens

Trigger: model_on_device() for the tokenizer yields an object that fails isinstance(obj, PreTrainedTokenizerBase) — e.g. the tokenizer field points at a model checkpoint instead of a tokenizer record, or the loader returned an unexpected wrapper.

Common situations: Wiring a raw model into the tokenizer field; tokenizer files missing so the loader fell back to a default object; transformers version changes altering tokenizer classes; corrupted tokenizer record in the DB.

Related errors

• Expected PreTrainedTokenizerBase for tokenizer, got {type(to
• Expected PreTrainedTokenizerBase for Gemma tokenizer, got {t
• Expected PreTrainedModel for Gemma encoder, got {type(gemma_
• Expected PreTrainedTokenizerBase for Gemma tokenizer, got {t
• Expected PreTrainedTokenizerBase for tokenizer, got {type(to

AI-assisted analysis of invoke-ai/InvokeAI@0b6a024f2f (2026-08-29). Data as JSON: /api/errors/456445c55d3d985f. Report an issue: GitHub.