sgl-project/sglang · error · ValueError

Unknown image_vae_encoding_position

Error message

Unknown image_vae_encoding_position: {image_vae_encoding_position}

What it means

This error is thrown by add_standard_ti2v_stages in the multimodal generation pipeline when the image_vae_encoding_position argument is neither the built-in 'before_timestep' position nor a value that triggers the dedicated image-VAE stage registration. It signals an unrecognized enum-like string in the text/image-to-video (TI2V) stage composition config. The library validates this setting because the VAE encoding placement determines where the image condition is injected into the denoising pipeline.

Solutions

1. Check the exact spelling and casing of image_vae_encoding_position against the values handled in composed_pipeline_base.py around the raise (the image_vae stage branch and 'before_timestep')
2. Set the argument to 'before_timestep' (the no-extra-stage default) or one of the supported stage positions
3. If you need a new position, extend the if/elif chain in add_standard_ti2v_stages before the raise

Example fix

# before
pipeline.add_standard_ti2v_stages(image_vae_encoding_position="before_timesteps")
# after
pipeline.add_standard_ti2v_stages(image_vae_encoding_position="before_timestep")

ALLOWED = {"before_timestep"}  # plus the image_vae stage branch values in add_standard_ti2v_stages
if pos not in ALLOWED:
    raise ConfigError(f"unsupported image_vae_encoding_position: {pos!r}; allowed: {sorted(ALLOWED)}")
pipeline.add_standard_ti2v_stages(image_vae_encoding_position=pos)

def is_valid_vae_encoding_position(pos: str) -> bool:
    return isinstance(pos, str) and pos in {"before_timestep"}

When it happens

Trigger: Calling create_pipeline_stages / add_standard_ti2v_stages with image_vae_encoding_position set to a string other than the supported values (the special-cased stage key or 'before_timestep'), e.g. a typo like 'before_timesteps' or 'after_timestep', or passing None/non-string when the pipeline expects a known position.

Common situations: Typos in pipeline config YAML/CLI, copying a position name from a different framework version where new positions were added/renamed, or upgrading/downgrading sglang where supported values changed.

Understand the failure class

Background: "Invalid value" and "allowed values are" config errors: what your library rejected and how to fix it — this error's family across 41 libraries.

Related errors

• Attention backend ' ' is not supported by this attention…
• audio_cap must be non-negative, got
• {error_msg}
• Expected grouped outputs, got
• Grouped pipeline returned fewer outputs than requests.

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/dda81ceb07ddcfe1. Report an issue: GitHub.

Appendix: source

Thrown at python/sglang/multimodal_gen/runtime/pipelines_core/composed_pipeline_base.py:983

                )
            )

        self.add_standard_latent_preparation_stage()
        self.add_standard_timestep_preparation_stage(
            prepare_extra_kwargs=prepare_extra_timestep_kwargs
        )
        if image_vae_encoding_position == "after_latent":
            self.add_stage(
                ImageVAEEncodingStage(
                    vae=self.get_module(image_vae_key),
                    **{
                        "component_name": image_vae_key,
                        **(image_vae_stage_kwargs or {}),
                    },
                )
            )
        elif image_vae_encoding_position != "before_timestep":
            raise ValueError(
                f"Unknown image_vae_encoding_position: {image_vae_encoding_position}"
            )

        if denoising_stage_factory is None:
            self.add_standard_denoising_stage()
        else:
            self.add_stage_factory(
                RoleType.DENOISER,
                denoising_stage_factory,
                denoising_stage_name,
            )

        self.add_standard_decoding_stage()
        return self

    # TODO(will): don't hardcode no_grad
    @torch.no_grad()
    def forward(

View on GitHub (pinned to 0132848349)