稀疏cpu模型训练优化和完善
Created by: jacquesqiao
异步训练性能优化
- Add async ssa graph executor https://github.com/PaddlePaddle/Paddle/pull/15409
- Optimize cpp reader https://github.com/PaddlePaddle/Paddle/pull/14731
- Add some debug tool https://github.com/PaddlePaddle/Paddle/pull/15471
- optimize profiler https://github.com/PaddlePaddle/Paddle/pull/14942
- 多cpu reduce模式不需要broadcast参数 https://github.com/PaddlePaddle/Paddle/pull/15661
- optimize concurrent index of selected rows for distribute lookup table https://github.com/PaddlePaddle/Paddle/pull/15698
稀疏参数训练功能完善
- 修复超大规模稀疏参数下merge_ids错误，会导致训练结果出现nan https://github.com/PaddlePaddle/Paddle/pull/15503 @seiriosPlus
- 解决训练会hang住的问题 @seiriosPlus https://github.com/PaddlePaddle/Paddle/pull/15503 一并修复了
- emb支持hash id的功能 @seiriosPlus
- emb支持动态增长的功能 @seiriosPlus
- update load persistables for increment @seiriosPlus https://github.com/PaddlePaddle/Paddle/pull/15576
文档：
- [] 超大规模稀疏参数(DistributedLookupTable)的本地预测+增量使用 https://github.com/PaddlePaddle/Paddle/issues/16360