Function build_mlp

src/framework/modules/attention/qwen_decoder.cpp:188–204

        v_heads.tensor,
        attention_mask.tensor,
        1.0F / std::sqrt(static_cast<float>(dim)),
        0.0F,
        0.0F);
    ggml_flash_attn_ext_set_prec(flash, precision);
    return core::wrap_tensor(
        flash,
        core::TensorShape::from_dims({q_contiguous.shape.dims[0], q_contiguous.shape.dims[2], q_contiguous.shape.dims[1], dim}),
        GGML_TYPE_F32);
}

core::TensorValue cache_view(
    core::ModuleBuildContext & ctx,
    const core::TensorValue & cache,
    int64_t start,
    int64_t steps,
    int64_t heads,
    int64_t dim) {
    if (start < 0 || steps <= 0 || start + steps > cache.shape.dims[1]) {
        throw std::runtime_error("Qwen decoder cache view range is invalid");
    }