huggingface/transformers · error · ImportError

Missing dependencies for serving. Install with `pip install

Error message

Missing dependencies for serving. Install with `pip install transformers[serving]`

What it means

The `transformers serve` command implements an OpenAI-compatible HTTP server using uvicorn and friends, which live behind the optional 'serving' extra. serve() checks is_serve_available() first and raises ImportError with the pip extra hint when the dependencies are missing, before any import of uvicorn happens.

Source

Thrown at src/transformers/cli/serve.py:109

        ] = None,
        cb_max_memory_percent: Annotated[
            float | None, typer.Option(help="Max GPU memory fraction for KV cache (0.0-1.0).")
        ] = None,
        cb_use_cuda_graph: Annotated[
            bool | None, typer.Option(help="Enable CUDA graphs for continuous batching.")
        ] = None,
        # Server options
        host: Annotated[str, typer.Option(help="Server listen address.")] = "localhost",
        port: Annotated[int, typer.Option(help="Server listen port.")] = 8000,
        enable_cors: Annotated[bool, typer.Option(help="Enable permissive CORS.")] = False,
        log_level: Annotated[str, typer.Option(help="Logging level (e.g. 'info', 'warning').")] = "warning",
        default_seed: Annotated[int | None, typer.Option(help="Default torch seed.")] = None,
        non_blocking: Annotated[
            bool, typer.Option(hidden=True, help="Run server in a background thread. Used by tests.")
        ] = False,
    ) -> None:
        if not is_serve_available():
            raise ImportError("Missing dependencies for serving. Install with `pip install transformers[serving]`")

        import uvicorn

        from .serving.chat_completion import ChatCompletionHandler
        from .serving.completion import CompletionHandler
        from .serving.model_manager import ModelManager
        from .serving.response import ResponseHandler
        from .serving.server import build_server
        from .serving.transcription import TranscriptionHandler
        from .serving.utils import GenerationState

        # Seed
        if default_seed is not None:
            set_torch_seed(default_seed)

        # Logging
        transformers_logger = logging.get_logger("transformers")
        transformers_logger.setLevel(logging.log_levels[log_level.lower()])

View on GitHub (pinned to a597f97485)

Solutions

1. pip install "transformers[serving]"
2. Or add uvicorn/fastapi (plus sse-starlette etc.) manually: pip install uvicorn fastapi
3. Rebuild deployment images with the serving extra included

Example fix

# before
pip install transformers
transformers serve ...  # ImportError

# after
pip install "transformers[serving]"
transformers serve ...

try:
    import uvicorn, fastapi  # noqa
except ImportError:
    raise SystemExit("pip install 'transformers[serving]' before running serve")

When it happens

Trigger: Running `transformers serve` in an install without the serving extras (plain `pip install transformers`); trimmed Docker images; lock files that stripped optional groups; partial upgrade that removed fastapi/uvicorn.

Common situations: Fresh venvs with base transformers only; CI images minimizing size; users upgrading transformers but not the extras; deploying to slim containers.

Related errors

• You need to install rich to use the chat interface. (`pip in
• You need to install `libcst` to run this command -> `pip ins
• Unknown error
• No server currently running on {url}. To run a local server,
• Unsupported dtype: '{dtype}'. Must be 'auto' or a valid torc

AI-assisted analysis of huggingface/transformers@a597f97485 (2026-08-14). Data as JSON: /api/errors/15b9bf5d9b18289b. Report an issue: GitHub.