DeepRec icon indicating copy to clipboard operation
DeepRec copied to clipboard

Published 20 hours ago • verified publisher icon DeepRec-AI

How to use multi level embedding?

Open allenlz opened this issue 2 years ago • 4 comments

I read https://mp.weixin.qq.com/s/aEi6ooG9wDL-GXVWcGWRCw and found DeepRec supports multi level embedding, which can put feature in HBM or DRAM by its hotness. It sounds a very good feature.

Then I read https://deeprec.readthedocs.io/zh/latest/Embedding-Variable.html but I can not found how to use multi level embedding.

My question is:

  1. How to use multi level embedding?

  2. If document is missing recently, could you give some config name or variable name as a clue? Then I can find related source code by myself.

Thanks.

allenlz avatar Apr 08 '22 09:04 allenlz

Multi-level doc haven't been added. @candyzone pls add a user doc asap. U can check the tensorflow/core/kernels/embedding_variable_ops_test.cc, 'StorageConfig' how to setup different storage level.

liutongxuan avatar Apr 08 '22 10:04 liutongxuan

BTW, is HBM_DRAM fully implemented in current master? It seems will use LocklessHashMap by default, which is only for cpu.

allenlz avatar Apr 11 '22 07:04 allenlz

@allenlz here is document. https://deeprec.readthedocs.io/zh/latest/Multi-tier-Embedding-Storage.html

candyzone avatar Jun 15 '22 07:06 candyzone

BTW, is HBM_DRAM fully implemented in current master? It seems will use LocklessHashMap by default, which is only for cpu.

Yes, we'll commit HBM_DRAM this month.

liutongxuan avatar Jun 16 '22 12:06 liutongxuan