How do you select these config candidates for cutlass?

I see the code in trtllm profiler.

My question is , why do you choose only these configs to tune?

How do you select them ? why there is no config with k=32 or k = 16 or some other m/n values ?

Thank you !

// Note: The shapes are in the format MxNxK. The K shape of the runtime config MUST match the K shape
//       in the kernel layout details when doing weight only quantization.
enum class CutlassTileConfig
{
    // Signals that we should run heuristics do choose a config
    Undefined,  

    // Signals that we should run heuristics do choose a config
    ChooseWithHeuristic,  

    // SiMT config
    CtaShape128x128x8_WarpShape64x64x8,    // CUDA core

    // TensorCore configs CTA_N = 128, CTA_K = 64
    // Warp configs for M=16
    CtaShape16x128x64_WarpShape16x32x64,
    // Warp configs for M=32
    CtaShape32x128x64_WarpShape32x32x64,

    // Warp configs for M=64
    CtaShape64x128x64_WarpShape32x64x64,
    CtaShape64x64x128_WarpShape32x64x64,
    CtaShape64x128x64_WarpShape64x32x64,

    // Warp configs for M=128
    CtaShape128x64x64_WarpShape64x32x64,
    CtaShape128x128x64_WarpShape64x32x64,
    CtaShape128x128x64_WarpShape64x64x64,
    CtaShape128x128x64_WarpShape128x32x64,
    CtaShape128x256x64_WarpShape64x64x64,

    // Warp configs for M=256
    CtaShape256x128x64_WarpShape64x64x64,

    // TensorCore config CTA_N = 64, CTA_K = 128
    CtaShape128x64x128_WarpShape64x32x128,

    // TensorCore config CTA_N = 256, CTA_K = 64
    CtaShape16x256x64_WarpShape16x64x64,

    // TensorCore config CTA_N = 256, CTA_K = 128
    CtaShape16x256x128_WarpShape16x64x128

};

Start with the TRT-LLM profiler code containing CutlassTileConfig and trace where these candidates are enumerated or passed to tuning. Compare the documented MxNxK and weight-only-quantization K constraints with the selection logic. Done means documenting why these configurations are included and whether other M/N/K configurations are intentionally excluded.