sgl-project/sglang · critical · Exception

{output_batch.error}

Error message

{output_batch.error}

What it means

Raised in DiffusionGenerator.generate after _send_to_scheduler_and_wait_for_response returns an OutputBatch whose error field is set. This means the scheduler/worker side of the diffusion pipeline reported a failure (model load error, inference exception, OOM, etc.) and the error string is propagated verbatim to the caller.

Source

Thrown at python/sglang/multimodal_gen/runtime/entrypoints/diffusion_generator.py:321

        results: list[GenerationResult] = []
        total_start_time = time.perf_counter()
        global_output_index = 0

        for requests in request_groups:
            try:
                timer_prompt = [req.prompt for req in requests]
                logger.info("Processing %d grouped request(s)", len(requests))
                with ExitStack() as stack:
                    for req in requests:
                        stack.enter_context(trace_req(req.trace_ctx))
                    timer = stack.enter_context(
                        log_generation_timer(logger, timer_prompt)
                    )
                    output_batch = self._send_to_scheduler_and_wait_for_response(
                        requests
                    )
                    if output_batch.error:
                        raise Exception(f"{output_batch.error}")

                    if (
                        output_batch.output is None
                        and output_batch.output_file_paths is None
                    ):
                        logger.error("Received empty output from scheduler")
                        continue

                    if requests[0].save_output and requests[0].return_file_paths_only:
                        output_file_paths = output_batch.output_file_paths or []
                        self._validate_output_count(
                            len(output_file_paths), len(requests)
                        )
                        for idx, path in enumerate(output_file_paths):
                            req = requests[idx]
                            if req.data_type == DataType.VIDEO:
                                req.sampling_params.validate_video_final_outputs(
                                    [path], req

View on GitHub (pinned to 0132848349)

Solutions

1. Read the inner error string — it carries the actual scheduler-side cause
2. Check scheduler logs for the traceback preceding this exception
3. Reduce batch size, image resolution, or number of inference steps if OOM
4. Verify model path/config and that input images are valid and readable

try:
    out = generator.generate(prompt=p)
except Exception as e:
    logger.error("scheduler failure: %s", e)
    # inspect scheduler logs / reduce batch before retrying

When it happens

Trigger: Any generate() call where the scheduler returns a non-None output_batch.error — e.g. worker crash during denoising, CUDA OOM, invalid latent shape, or an upstream preprocessing failure in the multimodal pipeline.

Common situations: GPU OOM with large batch/image sizes; model weights or VAE missing or mismatched; corrupted input image; scheduler process restarted mid-request.

Related errors

• {response.error}
• action policy returned no output
• Expected {request_count} outputs, got {output_count} from sc
• {failure_msg}: {error_msg}
• Subclasses of BaseScheduler must define '{attr}' property

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/a5db87702355e410. Report an issue: GitHub.