Profiled DDL RAMP

Collection of profiled models used to estimate the disrtibuted training time for different Transformer Encoder models partiotioned using Megatron partitioning strategy, for different target losses该数据集由经性能剖析的模型集合构成，可用于估算在不同目标损失配置下，采用Megatron分区策略（Megatron partitioning strategy）进行分区的各类Transformer编码器（Transformer Encoder）模型的分布式训练时长。