Add Grouped GEMM for Mixed Dtype

[muhammad-tanvir-1211]

This PR adds Grouped GEMM support for mixed precision GEMM.

[jiyang1011]

LGTM

[taozha2]

can we refine the include/cutlass/gemm/collective/xe_array_mma_mixed_input.hpp and include/cutlass/gemm/collective/xe_mma_mixed_input.hpp together, i found they are many common code.

[jiyang1011]

can we refine the include/cutlass/gemm/collective/xe_array_mma_mixed_input.hpp and include/cutlass/gemm/collective/xe_mma_mixed_input.hpp together, i found they are many common code.

can we refine the include/cutlass/gemm/collective/xe_array_mma_mixed_input.hpp and include/cutlass/gemm/collective/xe_mma_mixed_input.hpp together, i found they are many common code.

I tried to figure out the difference and dispatch it to xe_mma_mixed_input.hpp. the biggest diff is to initialize the params : array mma must initial the tiled copy with individual tensor and update the group index, so it is not easy

[taozha2]

can we refine the include/cutlass/gemm/collective/xe_array_mma_mixed_input.hpp and include/cutlass/gemm/collective/xe_mma_mixed_input.hpp together, i found they are many common code.

can we refine the include/cutlass/gemm/collective/xe_array_mma_mixed_input.hpp and include/cutlass/gemm/collective/xe_mma_mixed_input.hpp together, i found they are many common code.

I tried to figure out the difference and dispatch it to xe_mma_mixed_input.hpp. the biggest diff is to initialize the params : array mma must initial the tiled copy with individual tensor and update the group index, so it is not easy

the quantization and operator(gemm main loop) is same which is the most important part of the implementation, can we make a base struct like xe_mma_mixed_dtype_base contains these common part, and your grouped mixed gemm inherit it？

[jiyang1011]

can we refine the include/cutlass/gemm/collective/xe_array_mma_mixed_input.hpp and include/cutlass/gemm/collective/xe_mma_mixed_input.hpp together, i found they are many common code.

can we refine the include/cutlass/gemm/collective/xe_array_mma_mixed_input.hpp and include/cutlass/gemm/collective/xe_mma_mixed_input.hpp together, i found they are many common code.

I tried to figure out the difference and dispatch it to xe_mma_mixed_input.hpp. the biggest diff is to initialize the params : array mma must initial the tiled copy with individual tensor and update the group index, so it is not easy

the quantization and operator(gemm main loop) is same which is the most important part of the implementation, can we make a base struct like xe_mma_mixed_dtype_base contains these common part, and your grouped mixed gemm inherit it？

True, But this method will involve a lot of files. I will provide another PR to deal with it