sgl-project/sglang · error · ValueError

Subclass {self.__class__.__name__} must define _supported_at

Error message

Subclass {self.__class__.__name__} must define _supported_attention_backends

What it means

Raised by the text encoder base __init__ when the subclass's supported_attention_backends (from _supported_attention_backends) is empty. Every text encoder must declare which attention backends it supports; omitting it is an implementation error caught at construction.

Source

Thrown at python/sglang/multimodal_gen/runtime/models/encoders/base.py:226

    ]
    _fsdp_shard_conditions: list = field(default_factory=lambda: [])
    # Methods that drive a forward pass without going through __call__. FSDP2
    # only unshards around the wrapped module's own forward, so anything the
    # shard conditions left in the root group stays sharded unless the entry
    # point is registered; loaders read this and register each name.
    _fsdp_forward_methods: tuple[str, ...] = ()
    _stacked_params_mapping: list[tuple[str, str, str]] = field(default_factory=list)
    _supported_attention_backends: set[AttentionBackendEnum] = (
        TextEncoderConfig()._supported_attention_backends
    )

    def __init__(self, config: TextEncoderConfig) -> None:
        super().__init__()
        self.config = config
        self._fsdp_shard_conditions = config.arch_config._fsdp_shard_conditions
        self._stacked_params_mapping = config.arch_config.stacked_params_mapping
        if not self.supported_attention_backends:
            raise ValueError(
                f"Subclass {self.__class__.__name__} must define _supported_attention_backends"
            )

    @abstractmethod
    def forward(
        self,
        input_ids: torch.Tensor | None,
        position_ids: torch.Tensor | None = None,
        attention_mask: torch.Tensor | None = None,
        inputs_embeds: torch.Tensor | None = None,
        output_hidden_states: bool | None = None,
        **kwargs,
    ) -> BaseEncoderOutput:
        pass

    @property
    def supported_attention_backends(self) -> set[AttentionBackendEnum]:
        return self._supported_attention_backends

View on GitHub (pinned to 0132848349)

Solutions

1. Define _supported_attention_backends = ['flashattention', 'fa3', ...] (non-empty) on your subclass
2. Copy the declaration pattern from an existing encoder subclass

Example fix

# before
class MyEncoder(BaseTextEncoder):
    _supported_attention_backends: list[str] = []
# after
class MyEncoder(BaseTextEncoder):
    _supported_attention_backends = ["flashattention", "triton_attn"]

assert getattr(MyEncoder, "supported_attention_backends", None), "declare _supported_attention_backends"

def declares_backends(cls) -> bool:
    return bool(getattr(cls, "_supported_attention_backends", None))

When it happens

Trigger: Subclassing the base text encoder without defining _supported_attention_backends (or defining it as an empty list) and instantiating the subclass.

Common situations: Adding a new text encoder model class and forgetting the backend declaration required by the base class contract.

Related errors

• flashinfer_sparse_mla supports only GLM DSA with FP8 KV cach
• GLM DSA with FP8 KV cache on NVIDIA SM120/SM121 supports onl
• MiniMax-H3 ring parallelism requires the FlashAttention back
• AITer backend does not have a metadata builder.
• AITer backend requires num_heads ({num_heads}) to be a multi

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/bc4f7d84be0d6c10. Report an issue: GitHub.