sgl-project/sglang · error · TypeError

JoyEchoPipeline requires JoyEchoPipelineConfig, got {type(co

Error message

JoyEchoPipeline requires JoyEchoPipelineConfig, got {type(config)}

What it means

Raised by JoyEchoPipeline.create_pipeline_stages when server_args.pipeline_config is not a JoyEchoPipelineConfig. The stage-construction phase (memory bank, multishot setup, LTX2 front stages) reads settings only present on the typed config, so mismatched configs fail fast.

Source

Thrown at python/sglang/multimodal_gen/runtime/pipelines/joy_echo_pipeline.py:54

    def _get_or_create_memory_bank(
        self, config: JoyEchoPipelineConfig
    ) -> PairedAudioVideoMemoryBank:
        if self._memory_bank is None:
            self._memory_bank = PairedAudioVideoMemoryBank(
                max_size=int(config.memory_max_size),
                num_fix_frames=int(config.memory_num_fix_frames),
            )
        return self._memory_bank

    def reset_memory_bank(self) -> None:
        if self._memory_bank is not None:
            self._memory_bank.memory.clear()

    def create_pipeline_stages(self, server_args: ServerArgs):
        config = server_args.pipeline_config
        if not isinstance(config, JoyEchoPipelineConfig):
            raise TypeError(
                f"JoyEchoPipeline requires JoyEchoPipelineConfig, got {type(config)}"
            )

        memory_bank = self._get_or_create_memory_bank(config)
        self.add_stage(JoyEchoMultishotSetupStage(pipeline=self))
        _add_ltx2_front_stages(self)
        self.add_stage(JoyEchoSigmaPreparationStage())
        self.add_standard_timestep_preparation_stage(
            prepare_extra_kwargs=[prepare_ltx2_mu]
        )
        self.add_stages(
            [
                LTX2AVLatentPreparationStage(
                    scheduler=self.get_module("scheduler"),
                    transformer=self.get_module("transformer"),
                    audio_vae=self.get_module("audio_vae"),
                ),
                LTX2ImageEncodingStage(

View on GitHub (pinned to 0132848349)

Solutions

1. Construct JoyEchoPipelineConfig with the memory/multishot settings and assign it to server_args.pipeline_config before stage creation
2. Verify the pipeline registry selected JoyEchoPipeline for your model (a wrong registry mapping can pass the wrong config)
3. Add an isinstance check in your own startup script to give a clearer error location

Example fix

// before
server_args.pipeline_config = PipelineConfig()  # generic
pipeline.create_pipeline_stages(server_args)

// after
server_args.pipeline_config = JoyEchoPipelineConfig(...)
pipeline.create_pipeline_stages(server_args)

from sglang.multimodal_gen.runtime.pipelines.joy_echo_pipeline import JoyEchoPipelineConfig
if not isinstance(server_args.pipeline_config, JoyEchoPipelineConfig):
    server_args.pipeline_config = JoyEchoPipelineConfig()

def is_joyecho_config(cfg: object) -> TypeGuard[JoyEchoPipelineConfig]:
    return isinstance(cfg, JoyEchoPipelineConfig)

When it happens

Trigger: Advancing a JoyEcho pipeline to stage construction with server_args.pipeline_config set to None or another pipeline's config class; running the server with a JoyEcho model path but generic pipeline args.

Common situations: Copy-pasted launch scripts from another multimodal pipeline; configs produced by a factory that returns the base PipelineConfig; version skew where JoyEchoPipelineConfig moved modules.

Related errors

• Expected Hunyuan3D2PipelineConfig, got {type(config)}
• Expected scheduler.sigmas to be a tensor for JoyEcho.
• Unknown serve backend {name!r}. Available values: {available
• Multiple distributions register serve backend {name!r}: {pro
• Failed to load serve backend {name!r} from {self._entry_poin

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/d89f27925a7d63eb. Report an issue: GitHub.