Mila::Dnn::Quant::Weight::WeightQuantPolicy Concept Reference

Concept definition

        template<typename T>
        concept WeightQuantPolicy = requires
    {
        { T::kIsQuantized  } -> std::convertible_to<bool>;
        {
            T::kStorageDtype
        } -> std::convertible_to<TensorDataType>;
        {
            T::kScaleDtype
        } -> std::convertible_to<TensorDataType>;
        {
            T::kPerChannel
        } -> std::convertible_to<bool>;
        // Bits of the PRIMARY weight tensor per logical element -- what Linear needs to
        // size the packed allocation. It states the fact the allocation used to infer from
        // kPerChannel, which only ever meant "nibble-packed" because every per-group policy
        // happened to be 4-bit. A policy whose format spills into a companion plane counts
        // only the primary tensor here; the companion carries its own extent.
        {
            T::kStorageBitsPerElement
        } -> std::convertible_to<int>;
    }