sgl-project/sglang · critical · ValueError

Missing config.sliding_window for Mellum sliding_attention l

Error message

Missing config.sliding_window for Mellum sliding_attention layer {layer_id}

What it means

Mellum routes sliding-window attention per layer via config.layer_types; when a layer is 'sliding_attention' the model reads the window size from config.sliding_window. If that field is absent (None), the model cannot size the attention window and raises. It intentionally ignores legacy use_sliding_window post-init side effects.

Source

Thrown at python/sglang/srt/models/mellum.py:368

        if layer_type not in ("sliding_attention", "full_attention"):
            raise ValueError(
                f"Unsupported layer_types[{layer_id}]={layer_type}; "
                "expected 'sliding_attention' or 'full_attention'"
            )

        rope_parameters = cfg.rope_parameters
        rope_params = rope_parameters.get(layer_type)
        if rope_params is None:
            raise ValueError(
                f"Missing rope_parameters[{layer_type}] for Mellum layer {layer_id}"
            )

        # Mellum routes SWA per-layer via layer_types. Preserve the configured
        # window regardless of legacy use_sliding_window post-init side effects.
        if layer_type == "sliding_attention":
            sliding_window_size = get_attention_sliding_window_size(config)
            if sliding_window_size is None:
                raise ValueError(
                    "Missing config.sliding_window for Mellum "
                    f"sliding_attention layer {layer_id}"
                )
        else:
            sliding_window_size = -1

        max_position_embeddings = cfg.max_position_embeddings
        head_dim = cfg.head_dim
        rms_norm_eps = cfg.rms_norm_eps
        attention_bias = cfg.attention_bias

        self.self_attn = MellumAttention(
            hidden_size=self.hidden_size,
            num_heads=cfg.num_attention_heads,
            num_kv_heads=cfg.num_key_value_heads,
            layer_id=layer_id,
            start_layer=start_layer,
            rope_params=rope_params,

View on GitHub (pinned to 0132848349)

Solutions

1. Set an integer config.sliding_window in config.json (e.g. 8192) matching the checkpoint's window
2. Verify no post-init code sets use_sliding_window=False and nulls sliding_window before the model builds
3. If the checkpoint truly has no SWA, change that layer's layer_types entry to 'full_attention'

Example fix

// before
"layer_types": ["sliding_attention"], "sliding_window": null
// after
"layer_types": ["sliding_attention"], "sliding_window": 8192

cfg = AutoConfig.from_pretrained(path)
if "sliding_attention" in (cfg.layer_types or []):
    assert getattr(cfg, "sliding_window", None), "config.sliding_window must be set for SWA layers"

def swa_config_ok(cfg) -> bool:
    return "sliding_attention" not in (cfg.layer_types or []) or isinstance(getattr(cfg, "sliding_window", None), int)

When it happens

Trigger: Loading a Mellum config where layer_types contains 'sliding_attention' but sliding_window is null/missing in config.json; converting a checkpoint that stored the window under a different key (e.g. sliding_window in rope_parameters or per-layer overrides).

Common situations: Hand-written or stripped config.json during conversion; HF configs where sliding_window was set to None post-init; mixing config files across Mellum revisions.

Related errors

• Missing rope_parameters[{layer_type}] for Mellum layer {laye
• Expected len(mlp_layer_types) == num_hidden_layers, got {len
• Unsupported mlp_layer_types[{lid}]={mlp_type}; expected 'spa
• Sparse MLP requested but num_experts <= 0 in Mellum config
• bad compress_ratio {compress_ratio}

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/079357af79d6fc83. Report an issue: GitHub.