sgl-project/sglang · error · ValueError

layer_types contains unknown entries {bad}; expected only 'f

Error message

layer_types contains unknown entries {bad}; expected only 'full_attention' or 'sliding_attention'

What it means

Each entry of an explicitly provided layer_types list must be exactly 'full_attention' or 'sliding_attention'. Any other string (typos, 'sliding', 'local', 'full') or non-string value fails the set-difference check and raises this error.

Source

Thrown at python/sglang/srt/hardware_backend/mlx/models/muse_glimmer_mlx.py:305

        # NoPE layers are the full-attention layers; the rest slide.
        derived_layer_types = [
            "full_attention" if rope_flag == 0 else "sliding_attention"
            for rope_flag in self.no_rope_layers
        ]
        if self.layer_types is None:
            self.layer_types = derived_layer_types
        else:
            if len(self.layer_types) != self.num_hidden_layers:
                raise ValueError(
                    f"layer_types has {len(self.layer_types)} entries but "
                    f"num_hidden_layers is {self.num_hidden_layers}"
                )
            bad = sorted(
                set(self.layer_types) - {"full_attention", "sliding_attention"}
            )
            if bad:
                raise ValueError(
                    f"layer_types contains unknown entries {bad}; expected only "
                    "'full_attention' or 'sliding_attention'"
                )
            if self.layer_types != derived_layer_types:
                mismatches = [
                    i
                    for i, (got, want) in enumerate(
                        zip(self.layer_types, derived_layer_types)
                    )
                    if got != want
                ]
                raise ValueError(
                    "layer_types disagrees with no_rope_layers (NoPE layers "
                    "must be the full_attention layers) at layer indices "
                    f"{mismatches}"
                )

        if self.muse_glimmer_mlx_format is not None and (

View on GitHub (pinned to 0132848349)

Solutions

1. Replace every entry with exactly 'full_attention' or 'sliding_attention'
2. Set layer_types to null and let it be derived from no_rope_layers
3. If porting from another framework, map its vocabulary to these two strings programmatically

Example fix

// before
"layer_types": ["full", "sliding", "sliding"]
// after
"layer_types": ["full_attention", "sliding_attention", "sliding_attention"]

VALID = {"full_attention", "sliding_attention"}
assert set(cfg.get("layer_types") or []) <= VALID

def valid_layer_types(cfg: dict) -> bool:
    return set(cfg.get("layer_types") or []) <= {"full_attention", "sliding_attention"}

When it happens

Trigger: Providing layer_types with misspelled or differently-named attention type strings, e.g. 'full' instead of 'full_attention' or 'sliding_window' instead of 'sliding_attention'.

Common situations: Porting configs from other codebases (vLLM/transformers) that use different attention-type vocabulary, hand-writing layer_types, or LLM-assisted config generation introducing inconsistent naming.

Related errors

• layer_types has {len(self.layer_types)} entries but num_hidd
• layer_types disagrees with no_rope_layers (NoPE layers must
• bad compress_ratio {compress_ratio}
• The requested FlashAttention forward configuration exceeds S
• flashinfer_sparse_mla supports only GLM DSA with FP8 KV cach

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/8b367c2d8efef3cc. Report an issue: GitHub.