sgl-project/sglang · error · ValueError

Either `prompt` or `prompt_embeds` must be provided

Error message

Either `prompt` or `prompt_embeds` must be provided

What it means

The input-validation stage requires a textual prompt or precomputed prompt_embeds for every request except the I2M (image-to-motion) task, which can be image-only. If batch.prompt and batch.prompt_embeds are both None on a non-I2M task, generation cannot proceed because the text conditioner has no input.

Source

Thrown at python/sglang/multimodal_gen/runtime/pipelines_core/stages/input_validation.py:325

        self._generate_seeds(batch, server_args)

        if (
            server_args.pipeline_config.task_type == ModelTaskType.I2M
            and batch.num_inference_steps is None
            and hasattr(server_args.pipeline_config, "shape_num_inference_steps")
        ):
            batch.num_inference_steps = (
                server_args.pipeline_config.shape_num_inference_steps
            )

        # Ensure prompt is properly formatted (I2M can be image-only)
        if (
            server_args.pipeline_config.task_type != ModelTaskType.I2M
            and batch.prompt is None
            and batch.prompt_embeds is None
        ):
            raise ValueError("Either `prompt` or `prompt_embeds` must be provided")

        # Ensure negative prompt is properly formatted if using classifier-free guidance
        if (
            batch.do_classifier_free_guidance
            and batch.negative_prompt is None
            and batch.negative_prompt_embeds is None
        ):
            raise ValueError(
                "For classifier-free guidance, either `negative_prompt` or "
                "`negative_prompt_embeds` must be provided"
            )

        # Validate number of inference steps
        if batch.num_inference_steps <= 0:
            raise ValueError(
                f"Number of inference steps must be positive, but got {batch.num_inference_steps}"
            )

View on GitHub (pinned to 0132848349)

Solutions

1. Provide a non-None prompt (even a short caption) for non-I2M tasks
2. Or pass precomputed prompt_embeds if you encode prompts yourself
3. If the request really is image-only motion generation, set server_args.pipeline_config.task_type to ModelTaskType.I2M

Example fix

// before
out = pipe(prompt=None, image=img, task="i2v")

// after
out = pipe(prompt="a wave crashing", image=img, task="i2v")

if task_type != ModelTaskType.I2M:
    assert batch.prompt is not None or batch.prompt_embeds is not None, "non-I2M tasks need a prompt"

def request_has_text_input(batch, is_i2m: bool) -> bool:
    return is_i2m or batch.prompt is not None or batch.prompt_embeds is not None

When it happens

Trigger: Calling a T2V/T2I/I2V task with prompt=None and no prompt_embeds (e.g. relying on the image alone), or constructing a batch manually and forgetting to populate the prompt fields.

Common situations: Building image-to-video requests and assuming the image suffices (only valid for task_type == I2M); passing prompt as an empty string handled as None upstream; programmatic batch construction that skips prompt; disabling text conditioning flags while still using a text-conditioned task.

Related errors

• unsupported input for causal Conv3D cat/pad CUDA
• unsupported input for usp_merge_heads CUDA
• unsupported input for modulate_scale_shift CUDA
• unsupported input for LTX2 QKNorm split-RoPE CUDA
• fl2va requires first_frame, last_frame, or both

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/a3711a206b067fac. Report an issue: GitHub.