sgl-project/sglang · error · RuntimeError

Failed to generate MiniMax-H3 video: {str(e)}

Error message

Failed to generate MiniMax-H3 video: {str(e)}

What it means

RuntimeError wrapping any exception from sgld_client.generate_video() in the ComfyUI SGLDiffusion MiniMax-H3 node. The underlying cause (network error, server 4xx/5xx, invalid params) is included in the message; this is the node's single failure surface for video generation.

Source

Thrown at python/sglang/multimodal_gen/apps/ComfyUI_SGLDiffusion/nodes.py:780

            "audio_flow_shift": audio_flow_shift,
        }

        request_params = {
            "prompt": positive_prompt,
            "seconds": int(duration_seconds),
            "num_inference_steps": steps,
            "output_path": folder_paths.get_temp_directory(),
            "extra_fields": extra_fields,
        }
        if negative_prompt:
            request_params["negative_prompt"] = negative_prompt
        if seed >= 0:
            request_params["seed"] = seed

        try:
            response = sgld_client.generate_video(**request_params)
        except Exception as e:
            raise RuntimeError(f"Failed to generate MiniMax-H3 video: {str(e)}")

        video_path = response.get("file_path", "")
        # H3 aligns the canvas server-side, so the resolved size is only known
        # from the response; short_edge and aspect_ratio cannot reconstruct it
        resolved_size = response.get("size", "")
        if resolved_size:
            width, height = (int(v) for v in resolved_size.split("x"))
        else:
            width = height = short_edge
        video = convert_video_to_comfy_video(video_path, height, width)

        return (video, video_path)


class SGLDiffusionServerSetLora:
    """Node to set LoRA adapter for SGLang Diffusion server."""

    @classmethod

View on GitHub (pinned to 0132848349)

Solutions

1. Check the wrapped cause in the message: connection errors mean the server is down or URL is wrong — start/reconnect the SGLDiffusionServerModel node
2. Inspect server logs for the root cause (OOM, param validation)
3. Update the ComfyUI node pack and sglang to matching versions
4. Retry with smaller resolution/frame count if the server OOMed

assert sgld_client.health() == 'ok' before generating

try:
    out = node.generate(...)
except RuntimeError as e:
    log.error('H3 generation failed: %s', e)
    fallback_to_previous_checkpoint_or_retry()

When it happens

Trigger: Any exception thrown by the SGLang Diffusion server HTTP call: server down/unreachable, invalid request params, OOM or generation failure server-side, timeout.

Common situations: Server not started or wrong URL in the server node; server crashed mid-generation; invalid params (bad seed, unsupported resolution/fps); version mismatch between node and server API.

Related errors

• Video generation failed: {error_msg}
• Lost connection to server after {consecutive_errors} consecu
• Network error after {consecutive_errors} consecutive failure
• Video generation timed out after {max_wait_time} seconds
• Failed to generate video: {str(e)}

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/5242278921618bc8. Report an issue: GitHub.