ClikaRT::ops::qmoe_woq

function

qmoe_woq()

Tensor qmoe_woq(
    Tensor input,
    Tensor router_logits,
    QTensor fc1_experts,
    QTensor fc2_experts,
    std::int64_t top_k,
    OptionalTensor fc1_bias = {},
    OptionalTensor fc2_bias = {},
    QTensor fc3_experts = {},
    OptionalTensor fc3_bias = {},
    OptionalTensor e_score_correction_bias = {},
    OptionalTensor router_weights = {},
    std::optional<MoeRouting> routing_mode = std::nullopt,
    std::optional<bool> renormalize = std::nullopt,
    std::optional<std::int64_t> n_group = std::nullopt,
    std::optional<std::int64_t> topk_group = std::nullopt,
    std::optional<double> routed_scaling_factor = std::nullopt,
    std::optional<double> sparse_mixer_eps = std::nullopt,
    std::optional<bool> apply_router_weight_on_input = std::nullopt,
    std::optional<Activation> activation = std::nullopt,
    std::optional<SwigluFusion> swiglu_fusion = std::nullopt,
    std::optional<double> swiglu_alpha = std::nullopt,
    std::optional<double> swiglu_beta = std::nullopt,
    std::optional<double> swiglu_limit = std::nullopt,
    std::optional<GeluMode> gelu_mode = std::nullopt,
    OptionalTensor shared_output = {}
)

Declared in ClikaRT/compute/q_tensor.h, line 1243