Checkpoint size问题

我设置了个以embedding table为主的模型，设置ps_memory_m=64G，并且调整模型大小，使的placement正好占满memory(
再大一点就会报cannot placement的错误)。然后使用saver take了checkpoint，但是发现每个server的checkpoint大小都只有20G左右。请问为什么checkpoint的大小相比ps_memory_m要小这么多？XDL里什么部分使用了大量的内存，但是却不需要写到checkpoint里？

Start by tracing the saver checkpoint path and the memory accounting behind ps_memory_m and placement for the embedding-table model. Compare the memory reported for each server with the data actually serialized in its checkpoint; done means explaining the difference and identifying which memory is excluded from checkpoint output.

Written by the indexing model from the issue text.