sgl-project/sglang · error · RuntimeError

action policy returned no output

Error message

action policy returned no output

What it means

generate_action got a successful (non-error) response from the scheduler but output_batch.output is None, meaning the action policy produced no output payload. This indicates an empty/aborted generation on the worker side rather than an explicitly reported error.

Source

Thrown at python/sglang/multimodal_gen/runtime/entrypoints/diffusion_generator.py:473

            self.server_args.model_path,
            server_args=self.server_args,
            **sampling_params_kwargs,
        )
        if sampling_params.data_type != DataType.ACTION:
            raise ValueError(
                f"generate_action requires an ACTION pipeline, got {sampling_params.data_type}"
            )

        req = prepare_request(
            server_args=self.server_args,
            sampling_params=sampling_params,
            external_trace_header=external_trace_header,
        )
        output_batch = self._send_to_scheduler_and_wait_for_response(req)
        if output_batch.error:
            raise RuntimeError(output_batch.error)
        if output_batch.output is None:
            raise RuntimeError("action policy returned no output")
        return output_batch.output[0]

    def _resolve_prompts(
        self,
        prompt: str | list[str] | None,
        prompt_path: str | None = None,
    ) -> list[str]:
        """Collect prompts from the argument or from a prompt file."""
        path = prompt_path or self.server_args.prompt_file_path
        if path is not None:
            if not os.path.exists(path):
                raise FileNotFoundError(f"Prompt text file not found: {path}")
            with open(path, encoding="utf-8") as f:
                prompts = [line.strip() for line in f if line.strip()]
            if not prompts:
                raise ValueError(f"No prompts found in file: {path}")
            logger.info("Found %d prompts in %s", len(prompts), path)
            return prompts

View on GitHub (pinned to 0132848349)

Solutions

1. Log and inspect the full OutputBatch (output, output_file_paths, error) to see which fields are populated
2. Verify the prompt/observation passed to generate_action is non-empty and well-formed
3. Check scheduler logs for early-return/abort paths on this request
4. Retry the call once — if it consistently returns None, file it as a pipeline bug with the batch dump

assert prompt and prompt.strip(), "observation/prompt must be non-empty"

def is_valid_action_prompt(prompt) -> bool:
    return isinstance(prompt, str) and bool(prompt.strip())

try:
    action = generator.generate_action(prompt=obs)
except RuntimeError as e:
    if "no output" in str(e):
        action = default_action  # fallback policy
    else:
        raise

When it happens

Trigger: generate_action() where the scheduler returns successfully but attaches no output — e.g. request aborted, empty batch, or a worker path that returns early without populating outputs.

Common situations: Empty or whitespace-only observation/prompt; request cancelled mid-flight; edge case in the action pipeline where no frames were generated.

Related errors

• action policy returned no output
• {output_batch.error}
• Expected {request_count} outputs, got {output_count} from sc
• Subclasses of BaseScheduler must define '{attr}' property
• denoising_strength must be positive

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/c75ae1ca5d870126. Report an issue: GitHub.