[QST] Why does awq write its own int3/int4 GEMM kernels instead of using CUTLASS

[SimpleTheoryOfTypes]

Just curious about AWQ's decision to implement its own custom int3/int4 GEMM kernels rather than using NVIDIA's CUTLASS library (such as using int4 CUTLASS GEMMs developed by FasterTransformer). Was this choice driven by performance limitations with CUTLASS, or did they perhaps encounter compatibility issues? Thanks in advance for the insights!