VeLoRA: Memory Efficient Training using Rank-1 Sub-Token Projections | Read Paper on Bytez