Class MoeConfig

tensorrt_llm/layers/moe.py:106–142

@dataclass
class MoeConfig:

    class ExpertScaleNormalizationMode(IntEnum):
        NONE = 0
        RENORMALIZE = 1
        SPARSE_MIXER = 2
        DEVICE_LIMITED = 3
        DEVICE_LIMITED_RENORM = 4

    num_experts: int = 0
    shared_expert_intermediate_size: int = 0

    top_k: int = 0
    normalization_mode: ExpertScaleNormalizationMode = ExpertScaleNormalizationMode.RENORMALIZE
    sparse_mixer_epsilon: float = 0.01
    tp_mode: int = 0

    device_limited_n_group: int = 0
    device_limited_topk_group: int = 0
    device_limited_routed_scaling_factor: float = 1.0

    def validate(self) -> "MoeConfig":
        if (self.num_experts == 0) != (self.top_k == 0):
            raise ValueError(
                "Both or neither MoeConfig's num_experts and top_k must be set to 0"
            )
        return self

    def has_moe(self) -> bool:
        return self.num_experts > 1

    @classmethod
    def from_dict(cls, config: dict):
        return cls(**config)

    def to_dict(self):
        return asdict(self)


def _moe_plugin(moe_config,