sgl-project/sglang · error · ValueError

Invalid VAE type: {self.vae_type}

Error message

Invalid VAE type: {self.vae_type}

What it means

mova_audio_dit's frequency precomputation only supports vae_type == "dac"; any other value raises. The RoPE freqs depend on the audio VAE layout, so unsupported VAEs are rejected explicitly.

Source

Thrown at python/sglang/multimodal_gen/runtime/models/dits/mova_audio_dit.py:193

        self.previous_e0_even = None
        self.previous_e0_odd = None
        self.previous_residual_even = None
        self.previous_residual_odd = None
        self.is_even = False
        self.should_calc_even = True
        self.should_calc_odd = True
        self.accumulated_rel_l1_distance_even = 0
        self.accumulated_rel_l1_distance_odd = 0
        self.__post_init__()

    def _init_freqs(self):
        if self.freqs is not None:
            return
        head_dim = self.dim // self.num_heads
        if self.vae_type == "dac":
            self.freqs = precompute_freqs_cis_1d(head_dim)
        else:
            raise ValueError(f"Invalid VAE type: {self.vae_type}")

    def patchify(
        self,
        x: torch.Tensor,
        control_camera_latents_input: Optional[torch.Tensor] = None,
    ):
        x = self.patch_embedding(x)
        grid_size = x.shape[2:]
        x = rearrange(x, "b c f -> b f c").contiguous()
        return x, grid_size  # x, grid_size: (f)

    def unpatchify(self, x: torch.Tensor, grid_size: tuple[int]):
        return rearrange(
            x, "b f (p c) -> b c (f p)", f=grid_size[0], p=self.patch_size[0]
        )

    def forward(
        self,

View on GitHub (pinned to 0132848349)

Solutions

1. Set vae_type to "dac" in the audio DiT config (only supported audio VAE)
2. Check that the audio model config is not accidentally reusing the video model's vae_type
3. Update the library if you need a different audio VAE — it is unimplemented

Example fix

# before
audio_cfg = dict(vae_type="mova")
# after
audio_cfg = dict(vae_type="dac")

assert audio_cfg["vae_type"] == "dac", audio_cfg["vae_type"]

def is_supported_audio_vae(cfg) -> bool:
    return cfg.get("vae_type") == "dac"

When it happens

Trigger: Instantiating the audio DiT with a config whose vae_type is not "dac" (e.g. a video VAE type string leaking into the audio model config).

Common situations: Sharing one config object between video and audio model construction; typo'd vae_type; new audio codec support not implemented.

Related errors

• CuteDSL masked MoE supports activation 'silu' (gated) or 're

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/2056bb2289dcc1c1. Report an issue: GitHub.