vllm-project/vllm · error · TypeError

{init} received a positional argument of type {arg_type}, bu

Error message

{init} received a positional argument of type {arg_type}, but no parameter of that type was found in the method signature. Please either annotate {init} or pass it as a keyword argument.

What it means

The decorator patches the module's __init__ to inject vllm_config/prefix. As a safety check it walks the positional *args against the old __init__'s parameter annotations: if an argument's runtime type does not match the annotated type of the parameter in the same position, it raises TypeError suggesting you annotate __init__ or pass the argument by keyword. This guards against silently binding vllm_config/prefix into the wrong slots.

Source

Thrown at vllm/compilation/decorators.py:374

        vllm_config: VllmConfig | None = None,
        prefix: str = "",
        **kwargs: Any,
    ) -> None:
        if vllm_config is None:
            vllm_config = get_current_vllm_config()

        # NOTE: to support multimodal models (such as encoder),
        # we may not have vllm_config so we may need to patch it
        sig = inspect.signature(old_init)
        # Check that any positional arguments match the old_init method signature
        annotations = [p.annotation for p in sig.parameters.values()]
        for arg, annotation in zip(args, annotations):
            if annotation is inspect._empty:
                continue
            if not isinstance(arg, annotation):
                init = f"'{type(self).__name__}.__init__'"
                arg_type = f"'{type(arg).__name__}'"
                raise TypeError(
                    f"{init} received a positional argument of type {arg_type}, "
                    "but no parameter of that type was found in the method signature. "
                    f"Please either annotate {init} or pass it as a keyword argument."
                )
        if "vllm_config" in sig.parameters:
            kwargs["vllm_config"] = vllm_config
        if "prefix" in sig.parameters:
            kwargs["prefix"] = prefix
        old_init(self, *args, **kwargs)

        self.vllm_config = vllm_config
        self.compilation_config = self.vllm_config.compilation_config
        enable_compile = enable_if is None or enable_if(vllm_config)
        # for CompilationMode.STOCK_TORCH_COMPILE , the upper level model runner
        # will handle the compilation, so we don't need to do anything here.
        self.do_not_compile = (
            self.compilation_config.mode
            in [CompilationMode.NONE, CompilationMode.STOCK_TORCH_COMPILE]

View on GitHub (pinned to c794754062)

Solutions

1. Pass arguments as keyword arguments when constructing the module: MyModule(config=..., prefix=...).
2. Add accurate type annotations to every parameter of the decorated class's __init__.
3. Verify the positional order at the call site matches the annotated signature exactly.

Example fix

# before
module = MyDecoderLayer(vllm_config, "model.layers.0")  # TypeError if annotations mismatch
# after
module = MyDecoderLayer(vllm_config=vllm_config, prefix="model.layers.0")

import inspect

def bind_positional_safely(cls, args: tuple) -> None:
    params = list(inspect.signature(cls.__init__).parameters.values())[1:]
    for arg, p in zip(args, params):
        if p.annotation is not inspect._empty and not isinstance(arg, p.annotation):
            raise TypeError(f'{arg!r} does not match annotation {p.annotation}; pass by keyword')

try:
    layer = MyLayer(cfg, 'model.layers.0')
except TypeError as e:
    if 'pass it as a keyword argument' in str(e):
        layer = MyLayer(vllm_config=cfg, prefix='model.layers.0')
    else:
        raise

When it happens

Trigger: Instantiating a @support_torch_compile-decorated model with positional arguments whose order or type does not match annotated parameters — e.g. MyModule(config, 'prefix') where the second parameter is annotated torch.Tensor but receives a str.

Common situations: Custom model __init__ signatures reordered during refactoring while call sites kept the old positional order; __init__ parameters partially annotated; encoder/multimodal models whose vllm_config is optional so callers pass fewer positional args.

Related errors

• decorated class should have a forward method.
• No dynamic dimensions found in the forward method of {cls}.
• Argument {k} not found in the forward method of {cls}
• vLLM failed to compile the model. The most likely reason for
• shape_id='{shape_id}' requires PyTorch >= 2.11.0

AI-assisted analysis of vllm-project/vllm@c794754062 (2026-08-14). Data as JSON: /api/errors/2881a7e42821e840. Report an issue: GitHub.