torchrec

torchrec copied to clipboard

Back out "Remove fused input dist modules from train pipeline"

1

Summary: Original commit changeset: f2c6133871d1 Original Phabricator Diff: D47033218 Reviewed By: jmswen, joshuadeng Differential Revision: D47637260

ktsiam

Following the instructions under [torchrec/inference](https://github.com/pytorch/torchrec/blob/main/examples/inference/README.md), I get a Docker build failure. I'm using fresh ToT. Build log is [here](https://github.com/pytorch/torchrec/files/11963173/build.log).

coppock

Summary: # context * An ImportError was reported ``` ImportError: cannot import name 'register_custom_op' from 'torchrec.modules.utils' ``` * it's due to previously introduced function at D56443608 Differential Revision: D56578871

TroyGarden

Summary: This adds num positive and num negative examples in addition to total num examples in the QPS computation for torch metrics. Reviewed By: wilson100hong Differential Revision: D56376508

drdarshan

Summary: Test for PT2 aot_eager compilation covering ShardedEBC TW/CW. Test already contains initialization for FPEBC, EC model type. pt2/utils.py helper method for KJT for pt2_tracing. Differential Revision: D56653234

IvanKobzarev

Summary: mean pooling tests were verbose and in some cases not catching all cases (twrw requires 4 gpus to properly test) we refactor the tests here to apply for all...

zainhuda

KJT tensor compute path for List[int] *per_key

8

Differential Revision: D55932356

IvanKobzarev

Reviewed By: connernilsen Differential Revision: D56626638

connernilsen

How to share embeddings of some features between two EmbeddingBagCollections？

6

In a two-tower retrieval model, it is essential to randomly sample negative items. Typically, this means that the batch size for the item tower will be larger than that for...

tiankongdeguiji

Summary: Lookup embedding one batch ahead, alternating streams (blue / red pairing) {F1495167711} Differential Revision: D56508106

dstaay-fb