A different contribution was pointed out where a user produced a fused GEMM for int4, which happens to be effective for schooling with preset sequence lengths, giving the fastest Option.