sgl-project/sglang · critical · RuntimeError

Expected {request_count} outputs, got {output_count} from sc

Error message

Expected {request_count} outputs, got {output_count} from scheduler

What it means

DiffusionGenerator._validate_output_count checks that the number of results returned by the scheduler equals the number of requests sent. A mismatch means the scheduler dropped, duplicated, or mis-batched outputs — an internal consistency failure, not a user-input problem.

Source

Thrown at python/sglang/multimodal_gen/runtime/entrypoints/diffusion_generator.py:550

        else:
            size = (req.height, req.width, req.num_frames)
        return dict(
            prompt=req.prompt,
            size=size,
            generation_time=generation_time,
            peak_memory_mb=output_batch.peak_memory_mb,
            metrics=metrics.to_dict() if metrics else {},
            action=output_batch.action_pred,
            trajectory_latents=output_batch.trajectory_latents,
            trajectory_timesteps=output_batch.trajectory_timesteps,
            rollout_trajectory_data=output_batch.rollout_trajectory_data,
            trajectory_decoded=output_batch.trajectory_decoded,
        )

    @staticmethod
    def _validate_output_count(output_count: int, request_count: int) -> None:
        if output_count != request_count:
            raise RuntimeError(
                f"Expected {request_count} outputs, got {output_count} from scheduler"
            )

    def _send_to_scheduler_and_wait_for_response(self, batch: list[Req]) -> OutputBatch:
        """
        Sends a request to the scheduler and waits for a response.
        """
        return sync_scheduler_client.forward(batch)

    # LoRA
    def _send_lora_request(self, req: Any, success_msg: str, failure_msg: str):
        response = sync_scheduler_client.forward(req)
        if response.error is None:
            logger.info(success_msg)
            return response
        else:
            error_msg = response.error
            raise RuntimeError(f"{failure_msg}: {error_msg}")

View on GitHub (pinned to 0132848349)

Solutions

1. Retry the call once — transient scheduler hiccups can cause one-off mismatches
2. Update/align sglang versions between client and scheduler processes
3. Check scheduler logs for aborted or failed individual requests within the batch
4. If reproducible, reduce the batch to isolate which prompt triggers the mismatch and report it with logs

for attempt in range(2):
    try:
        return generator.generate(prompt=prompts)
    except RuntimeError as e:
        if "from scheduler" not in str(e) or attempt == 1:
            raise
        logger.warning("output count mismatch, retrying: %s", e)

When it happens

Trigger: generate() with N prompts where output_batch.output (or output_file_paths) has length != N; typically after a scheduler bug, a mixed output mode, or a partial failure where some requests produced files and others produced tensors.

Common situations: Scheduler version skew after a partial upgrade; a request in the batch aborted so its output is missing; mixing output modes (some requests writing files, some returning latents) so one list is shorter.

Related errors

• {output_batch.error}
• action policy returned no output
• Subclasses of BaseScheduler must define '{attr}' property
• denoising_strength must be positive
• Must pass a value for `mu` when `use_dynamic_shifting` is Tr

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/c9201f985134e228. Report an issue: GitHub.