[QST] why the implementation of f16xs8 mixed gemm is different between TRT-LLM and native cutlass mixed gemm example?

What is your question?
Dear cutlass team,

lets consider sm80 and f16s8, the example of f16s8 TN mixed gemm shown here is different from TRT-LLM implementation, specifically, to my knowledge, the TRT-LLM one added the dequantization scale, but the cutlass one did not. Then my questions are:

1. Is the performance or accuracy of TRT-LLM adding dequantization scale better than cutlass native one in LLM linear cases?
2. from here, I see the TRT-LLM one seems load operand B(s8) using LDS not LDSM, but I can't find the f16s8 LDS specialization in MmaTensorOpMultiplicandTileIterator, only find LDS specialization for TF32, which make me confused with the “LDS". Am I missing something?

Thanks your time!

cc @manishucsd @alexsamardzic @hwu36

Compare the native CUTLASS f16s8 TN mixed GEMM example linked from PR 1084 with TensorRT-LLM's fpA_intB_gemm_template.h and dq_mma_pipelined_finegrained.h. Trace the referenced MmaTensorOpMultiplicandTileIterator implementation for the sm80 path. Done means documenting the dequantization-scale and LDS/LDSM differences and answering their performance, accuracy, and specialization questions.