sgl-project/sglang · error · ValueError

action_mode is set but the loaded Cosmos3 checkpoint has no

Error message

action_mode is set but the loaded Cosmos3 checkpoint has no action modality (action_gen is False).

What it means

The request enables action generation via sampling_params.action_mode, but the loaded transformer has action_dim=None, i.e. the checkpoint lacks the action head. Action latents cannot be prepared without it, so the request is rejected.

Source

Thrown at python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/cosmos3.py:616

        if sound_duration > 0.0:
            if not getattr(self.transformer, "sound_gen", False):
                raise ValueError(
                    "sound generation was requested (sound_duration > 0) but the "
                    "loaded Cosmos3 checkpoint has no sound modality (sound_gen is "
                    "False)."
                )
            sound_latent_fps = self.transformer.sound_latent_fps
            sound_latent_frames = max(1, round(sound_duration * sound_latent_fps))
            sound_shape = (1, self.transformer.sound_dim, sound_latent_frames)
            batch.audio_latents = torch.randn(
                sound_shape, generator=generator, device=device, dtype=dtype
            )
            self.log_info(f"Prepared sound latents with shape {sound_shape}")

        action_mode = getattr(batch.sampling_params, "action_mode", None)
        if action_mode is not None:
            if getattr(self.transformer, "action_dim", None) is None:
                raise ValueError(
                    "action_mode is set but the loaded Cosmos3 checkpoint has no "
                    "action modality (action_gen is False)."
                )
            self._prepare_action_latents(batch, generator, device, dtype)
        return batch

    def component_uses(
        self, server_args: ServerArgs, stage_name: str | None = None
    ) -> list[ComponentUse]:
        return [ComponentUse(self._component_stage_name(stage_name), "vae")]

    @staticmethod
    def _resolve_domain_id(batch: Req) -> int:
        """Resolve action embodiment domain ID; required for action generation."""
        domain_id = getattr(batch.sampling_params, "domain_id", None)
        if domain_id is not None:
            domain_id = int(domain_id)
            if domain_id < 0:

View on GitHub (pinned to 0132848349)

Solutions

1. Remove action_mode from sampling params for non-action checkpoints
2. Load an action-capable Cosmos3 checkpoint (action_dim is set)
3. Check transformer.action_dim after load and gate action features on it

Example fix

# before
sp.action_mode = 'policy'  # checkpoint has no action head

# after
del sp.action_mode  # or load action-capable checkpoint

if sp.action_mode is not None and getattr(model.transformer, 'action_dim', None) is None:
    sp.action_mode = None  # checkpoint cannot do actions

def checkpoint_supports_action(transformer) -> bool:
    return getattr(transformer, 'action_dim', None) is not None

When it happens

Trigger: Setting action_mode (e.g. 'policy' or 'forward_dynamics') on a checkpoint whose transformer.action_dim attribute is None (non-action Cosmos3 variant).

Common situations: Mixing robotics/action workflows with a plain video-generation checkpoint; assuming all Cosmos3 builds include action heads.

Related errors

• sound generation was requested (sound_duration > 0) but the
• domain_id must be non-negative, got {domain_id}
• Unknown action domain name {domain_name!r}. Valid names: {so
• action must have shape [T, D], got {tuple(action.shape)}
• Cosmos3 action input accepts one image field; use a list or

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/bcbc99e464ee7434. Report an issue: GitHub.