We typically support whatever the underlying library supports. For int8 the support would come from cuBLASLt currently. I don't believe that or Cutlass supports mixed precision inputs, but I can check.