sgl-project/sglang · error · ValueError

Unsupported AVAE normalization_type={self.normalization_type

Error message

Unsupported AVAE normalization_type={self.normalization_type!r}.

What it means

_denormalize_latent supports 'none', tanh-based, and group-norm style normalization; anything else in normalization_type leaves latents un-denormalized, which would silently corrupt decode output, so an unrecognized value raises at decode time.

Source

Thrown at python/sglang/multimodal_gen/runtime/models/vaes/cosmos3_avae.py:212

        return self.hop_size

    def get_latent_num_samples(self, num_audio_samples: int) -> int:
        return int(num_audio_samples) // self.hop_size

    def get_audio_num_samples(self, num_latent_samples: int) -> int:
        return int(num_latent_samples) * self.hop_size

    def _denormalize_latent(self, latent: torch.Tensor) -> torch.Tensor:
        if self.normalization_type == "tanh":
            in_dtype = latent.dtype
            x = torch.clamp(
                latent.float() / self.tanh_output_scale,
                -self.tanh_clamp,
                self.tanh_clamp,
            )
            return (torch.atanh(x) * self.tanh_input_scale).to(in_dtype)
        if self.normalization_type != "none":
            raise ValueError(
                f"Unsupported AVAE normalization_type={self.normalization_type!r}."
            )
        return latent

    @torch.no_grad()
    def decode(self, latent: torch.Tensor) -> torch.Tensor:
        squeeze = latent.ndim == 2
        if squeeze:
            latent = latent.unsqueeze(0)
        decoder_dtype = next(self.decoder.parameters()).dtype
        decoder_device = next(self.decoder.parameters()).device
        z = self._denormalize_latent(latent.to(decoder_device)).to(decoder_dtype)
        audio = self.decoder(z).clamp(-1.0, 1.0).to(latent.dtype)
        return audio.squeeze(0) if squeeze else audio


EntryClass = Cosmos3AVAEAudioTokenizer

View on GitHub (pinned to 0132848349)

Solutions

1. Set normalization_type to a supported value ('none' or the tanh/group variants accepted by this class)
2. Check the exact supported set in this file's normalization parsing (just above, near line 195-210) and align spelling
3. Upgrade the library if the checkpoint needs a newer normalization type

Example fix

# before
config = {"normalization_type": "LayerNorm"}
# after
config = {"normalization_type": "none"}

SUPPORTED_NORMS = {"none"}  # plus tanh/group variants per this class
if config.get("normalization_type", "none") not in SUPPORTED_NORMS:
    config["normalization_type"] = "none"  # or fail loudly at load time

def is_supported_norm(t: str) -> bool:
    return t in {"none"}  # extend with the class's supported set

try:
    frames = avae.decode(latent)
except ValueError as e:
    if "normalization_type" in str(e):
        raise RuntimeError("bad AVAE config; fix normalization_type") from e
    raise

When it happens

Trigger: Decoding with a Cosmos3 AVAE whose config sets normalization_type to an unsupported string (typo like 'groupnorm ' or 'layer_norm') or a new scheme this version doesn't implement.

Common situations: Config typos or casing differences; porting a config from a newer Cosmos3 version that added a new normalization type; checkpoint config using an alias not handled by this code version.

Related errors

• Cosmos3 AVAE dec_strides product must equal hop_size: produc
• This browser cannot encode H.264 MP4
• H.264 encoder did not return MP4 decoder config
• This browser does not support gzip stream decoding
• delta payload size mismatch: expected ${expectedSize}, got $

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/56e1c2cfeaaba5e6. Report an issue: GitHub.