sgl-project/sglang · critical · ValueError

Unsupported attention type: {config.attention_type}

Error message

Unsupported attention type: {config.attention_type}

What it means

Bailing v3 layer construction dispatches on config.attention_type; unrecognized values (neither full nor the supported linear attention type) raise this error before the MoE layout is built.

Source

Thrown at python/sglang/srt/models/bailing_moe_v3.py:972

                    quant_config=quant_config,
                    layer_id=layer_id,
                    reduce_results=False,
                    prefix=add_prefix("attention", prefix),
                    alt_stream=alt_stream,
                    skip_rope=(
                        getattr(config, "use_mla_nope", False)
                        or config.qk_rope_head_dim == 0
                    ),
                )
            else:
                self.attention = BailingMoEAttention(
                    config,
                    quant_config=quant_config,
                    layer_id=self.layer_id,
                    prefix=prefix + ".attention",
                )
        else:
            raise ValueError(f"Unsupported attention type: {config.attention_type}")

        self.expert_num = config.num_experts
        self.hidden_size = config.hidden_size
        is_moe_layer = is_nextn or (
            not (self.expert_num == 1)
            and (self.layer_id >= config.first_k_dense_replace)
        )
        self.is_layer_sparse = is_moe_layer
        is_previous_moe_layer = not (self.expert_num == 1) and (
            self.layer_id - 1 >= config.first_k_dense_replace
        )
        is_next_layer_sparse = not (self.expert_num == 1) and (
            self.layer_id + 1 >= config.first_k_dense_replace
        )
        if enable_moe_dense_fully_dp():
            mlp_tp_rank, mlp_tp_size = 0, 1
        else:
            mlp_tp_rank, mlp_tp_size = None, None

View on GitHub (pinned to 0132848349)

Solutions

1. Upgrade SGLang to a release supporting this attention_type
2. Fix attention_type in config.json to a supported value ("full" or the model's linear attention type)

assert config.attention_type in {"full", "linear"}

When it happens

Trigger: Loading a bailing_moe_v3 config whose attention_type is unsupported by the installed SGLang version.

Common situations: Newer checkpoint revision with a new attention flavor; typo in attention_type; downgrade of SGLang.

Related errors

• Unsupported attention type: {config.attention_type}
• head_dim mismatch across layers for fused KV path: expected
• num_heads must be divisible by num_epi_subtiles
• num_heads // num_epi_subtiles must be divisible by 4 (FMA un
• sparse_attn_v4_paged_decode expects fp16/bf16 q, got {q.dtyp

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/7a68c92c0e66f640. Report an issue: GitHub.