composable_kernel Add custom type vector support

add vectorization support for custom types defined as non-native types or user-defined objects
add a test to check if native and non-native vector types take same expected space in memory

Jun 12 '24 19:06 geyyer

Added vector reshaping tests in dc7208c

Sep 17 '24 17:09 geyyer

When this PR could be merged?

Oct 07 '24 22:10 zjing14

Could you provide an example for custom data type for packed int4/fp4?

Oct 07 '24 22:10 zjing14

@zjing14 we would like to add a couple more tests and this PR should be good to go. We will add a GEMM example a little later. Are you interested in a regular FP4 or MXFP4?

Oct 07 '24 22:10 geyyer

We will add a GEMM example a little later. Are you interested in a regular FP4 or MXFP4?

We can start with mixed-precision GEMM w/ bf16xint4.

Oct 08 '24 19:10 zjing14

@zjing14, could you share more details: what kind of compute types are you looking for? Is it one of the int4 dequantized to bf16 type of kernels?

Oct 08 '24 20:10 geyyer

Yes, int4 dequantized to bf16.

Oct 09 '24 01:10 zjing14