DeLA v2 is faster, more memory efficient and stronger in expressivity. (Don't expect faster training though, there are many reasons: CPU bottleneck; Model scaled; CUDA kernels in fp32; prolonged steps ...