Lightning-AI/pytorch-lightning · error · Exception

Please, return your outputs as a dictionary. Found {output}

Error message

Please, return your outputs as a dictionary. Found {output}

What it means

Inside the spawned server, after deserializing the request body, servable_model.serve_step(**body) must return a dict so outputs can be matched with output serializers by key. If it returns a tensor, list, or tuple, a generic Exception is raised in the request handler.

Source

Thrown at src/lightning/pytorch/serve/servable_module_validator.py:170

        # Note: This isn't the original version, but a copy.
        servable_model.eval()

        @app.get("/ping")
        def ping() -> bool:
            return True

        @app.post("/serve")
        async def serve(payload: dict = Body(...)) -> dict[str, Any]:
            body = payload["body"]

            for key, deserializer in deserializers.items():
                body[key] = deserializer(body[key])

            with torch.no_grad():
                output = servable_model.serve_step(**body)

            if not isinstance(output, dict):
                raise Exception(f"Please, return your outputs as a dictionary. Found {output}")

            for key, serializer in serializers.items():
                output[key] = serializer(output[key])

            return output

        run(app, host=host, port=port, log_level="error")

View on GitHub (pinned to 9fed5c27d2)

Solutions

1. Change serve_step to return a dict whose keys match configure_serialization's output serializer names, e.g. {"output": result}
2. For multiple outputs return {"logits": ..., "probs": ...} with matching serializers

Example fix

# before
def serve_step(self, **kwargs):
    return self(kwargs["x"])
// after
def serve_step(self, **kwargs):
    return {"output": self(kwargs["x"])}

out = model.serve_step(**model.configure_payload()["body"])
assert isinstance(out, dict), f"serve_step must return a dict, got {type(out)}"

def returns_dict_serve_step(model, body) -> bool:
    with torch.no_grad():
        return isinstance(model.serve_step(**body), dict)

When it happens

Trigger: serve_step returns self(x) (a raw Tensor), a tuple of tensors, or None instead of a dict keyed by output names.

Common situations: Reusing predict_step code that returns a bare tensor; returning multiple outputs as a tuple and expecting positional serialization.

Related errors

• The `configure_payload` method needs to be overridden.
• The `configure_serialization` method needs to be overridden.
• The `serve_step` method needs to be overridden.
• The server didn't start within {self.timeout} seconds.
• Your provided payload {payload} should have a field named "b

AI-assisted analysis of Lightning-AI/pytorch-lightning@9fed5c27d2 (2026-08-28). Data as JSON: /api/errors/323450cd6beba7c2. Report an issue: GitHub.