sgl-project/sglang · error · ValueError

unknown norm_type {norm_type}

Error message

unknown norm_type {norm_type}

What it means

Raised by the conditioning embedding projection module in glm_image.py when norm_type is not 'layer_norm' or 'rms_norm'. The __init__ chooses between nn.LayerNorm and nn.RMSNorm based on this string; anything else is rejected. This guards against silently running without normalization, which would corrupt checkpoint loading and outputs.

Source

Thrown at python/sglang/multimodal_gen/runtime/models/dits/glm_image.py:886

        embedding_dim: int,
        conditioning_embedding_dim: int,
        elementwise_affine: bool = True,
        eps: float = 1e-5,
        bias: bool = True,
        norm_type: str = "layer_norm",
    ):
        super().__init__()
        self.linear = nn.Linear(
            conditioning_embedding_dim, embedding_dim * 2, bias=bias
        )
        if norm_type == "layer_norm":
            self.norm = nn.LayerNorm(embedding_dim, eps, elementwise_affine, bias)
            # For now, don’t replace this with sglang’s LayerNorm
            # because the model doesn’t have this parameter and it will break model loading
        elif norm_type == "rms_norm":
            self.norm = nn.RMSNorm(embedding_dim, eps, elementwise_affine)
        else:
            raise ValueError(f"unknown norm_type {norm_type}")

    def forward(
        self, x: torch.Tensor, conditioning_embedding: torch.Tensor
    ) -> torch.Tensor:
        # *** NO SiLU here ***
        emb = self.linear(conditioning_embedding.to(x.dtype))
        scale, shift = torch.chunk(emb, 2, dim=1)
        if is_plain_layer_norm(self.norm, x.shape[-1]):
            return _glm_ln_modulate(self.norm, x, scale, shift, x.dtype)
        x = self.norm(x) * (1 + scale)[:, None, :] + shift[:, None, :]
        return x


class GlmImageTransformer2DModel(CachableDiT, LayerwiseOffloadableModuleMixin):
    r"""
    Args:
        patch_size (`int`, defaults to `2`):
            The size of the patches to use in the patch embedding layer.

View on GitHub (pinned to 0132848349)

Solutions

1. Set norm_type to exactly 'layer_norm' or 'rms_norm' in the model/conditioning config
2. Verify against the checkpoint's original config which normalization the conditioning projection was trained with
3. Add a config sanitizer that maps alternative spellings to the two accepted values before model construction

Example fix

# before
{"norm_type": "none"}

# after
{"norm_type": "layer_norm"}

assert cfg['norm_type'] in ('layer_norm', 'rms_norm'), f"norm_type must be layer_norm|rms_norm, got {cfg['norm_type']!r}"

def is_valid_norm_type(v: str) -> bool:
    return v in ('layer_norm', 'rms_norm')

When it happens

Trigger: Instantiating the conditioning module (e.g. via the GLM image DiT) with norm_type set to a string other than 'layer_norm' or 'rms_norm', such as 'none', 'group_norm', or a casing variant like 'LayerNorm'.

Common situations: Mismatched config keys after converting a checkpoint from another framework (e.g. diffusers-style 'norm_type' values) into this model's expected schema; typos in hand-written configs.

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

• unknown qk_norm: {qk_norm}. Should be one of None, 'layer_no
• Unknown history_scale_mode: {history_scale_mode}
• Hidden size {hidden_size} must be divisible by num_heads {nu
• Got {axes_dim} but expected positional dim {pe_dim}
• denoising_strength must be positive

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/c167ba660193ed75. Report an issue: GitHub.