Method build

src/framework/modules/primitive_modules.cpp:171–197

    std::array<int, core::kMaxTensorRank> axes = {0, 1, 2, 3};
    std::swap(axes[static_cast<size_t>(axis)], axes[input.shape.rank - 1]);
    const auto transposed = TransposeModule({axes, input.shape.rank}).build(ctx, input);
    const auto contiguous = tensor_layout::ensure_contiguous_layout_if_needed(ctx, transposed);
    auto reduced = core::wrap_tensor(
        fn(ctx.ggml, contiguous.tensor),
        reduced_shape(transposed.shape, static_cast<int>(transposed.shape.rank - 1)),
        GGML_TYPE_F32);
    return TransposeModule({axes, reduced.shape.rank}).build(ctx, reduced);
}

}  // namespace

const core::ModuleSchema & MatMulModule::schema() const noexcept {
    return static_schema();
}

core::TensorValue MatMulModule::build(
    core::ModuleBuildContext & ctx,
    const core::TensorValue & lhs,
    const core::TensorValue & rhs) const {
    if (ctx.ggml == nullptr) {
        throw std::runtime_error("ModuleBuildContext.ggml is null");
    }
    core::validate_rank_between(lhs, 2, core::kMaxTensorRank, "lhs");
    core::validate_rank_between(rhs, lhs.shape.rank, lhs.shape.rank, "rhs");

    const size_t rank = lhs.shape.rank;
    for (size_t i = 0; i + 2 < rank; ++i) {
        if (lhs.shape.dims[i] != rhs.shape.dims[i]) {
            throw std::runtime_error("MatMul batch dimensions must match");
        }