zylon-ai/private-gpt · error · ImportError

Transformers dependencies are not installed.

Error message

Transformers dependencies are not installed.

What it means

HuggingFaceTokenizer.from_pretrained imports transformers.AutoProcessor inside the method so that the rest of the module works without transformers installed. If the import fails it raises ImportError with a formatted missing-dependency message for 'Transformers'.

Source

Thrown at private_gpt/components/llm/tokenizers/huggingface.py:69

        **kwargs: Any,
    ) -> "HuggingFaceTokenizer":
        """Load tokenizer from pretrained model with intelligent caching.

        If the model is already cached locally, it will automatically use
        offline mode to avoid network calls.

        Args:
            model_id: Model identifier or local path
            local_files_only: Force offline mode (no downloads)
            cache_dir: Custom cache directory
            force_download: Force re-download even if cached
            trust_remote_code: Allow custom code from model repositories
            **kwargs: Additional arguments for AutoProcessor
        """
        try:
            from transformers import AutoProcessor  # ty:ignore[unresolved-import]
        except ImportError as e:
            raise ImportError(
                format_missing_dependency_message(
                    "Transformers",
                )
            ) from e

        try:
            is_multimodal = False
            processor = None
            loaded: Any = AutoProcessor.from_pretrained(
                pretrained_model_name_or_path=model_id,
                local_files_only=local_files_only,
                cache_dir=cache_dir,
                force_download=force_download,
                trust_remote_code=trust_remote_code,
                **kwargs,
            )

            # Extract tokenizer from multimodal processor if needed

View on GitHub (pinned to 4a030776a3)

Solutions

1. Install transformers (or the project extra that pulls it in, e.g. the local-LLM extra): uv sync --inexact --extra <local-llm-extra>.
2. Verify: python -c "from transformers import AutoProcessor".
3. If you only use remote APIs and don't need real tokenization, switch tokenizer_mode to 'estimator' or 'remote'.

Example fix

# before: tokenizer_mode='huggingface' without transformers installed
# after
uv sync --inexact --extra llm-huggingface  # or pip install transformers

def transformers_available() -> bool:
    try:
        from transformers import AutoProcessor  # noqa: F401
        return True
    except ImportError:
        return False

try:
    tok = HuggingFaceTokenizer.from_pretrained(model_id)
except ImportError as e:
    logger.warning('transformers missing; falling back to estimator tokenizer')
    tok = TokenizerRegistry.get_tokenizer('estimator', model_id=model_id)

When it happens

Trigger: Calling HuggingFaceTokenizer.from_pretrained (directly or via TokenizerRegistry modes 'huggingface'/'chat') in an environment lacking the transformers package.

Common situations: Slack/minimal installs that skipped the tokenizer extras; Docker images trimmed for remote-API-only deployments (no local models) where tokenizer_mode was later changed to 'huggingface'; CI environments with a reduced dependency set.

Related errors

• HuggingFaceTokenizer is not available with the given configu
• Mistral tokenizer dependencies are not installed. Install wi
• tiktoken dependencies are not installed. Install with `uv sy
• DB2 database query dependencies are not installed. Install w
• OpenAI embeddings dependencies are not installed. Install wi

AI-assisted analysis of zylon-ai/private-gpt@4a030776a3 (2026-08-15). Data as JSON: /api/errors/33022a952e797eb9. Report an issue: GitHub.