torchrec

torchrec copied to clipboard

Fix O(N^2) scaling in enumerator for large models

1

Summary: EmbeddingEnumerator.enumerate() is O(N^2) with respect to the number of tables. For large models (5000 tables), this is very noticeable. This diff fixes the N^2 problem by constructing our sharding_options...

henrylhtsang

fix test SilvertorchUvStandAlonePublishBuckTest

3

Summary: As titled Reviewed By: tissue3 Differential Revision: D51864844

gnahzg

Change KJT length_per_key from List[int] to Tensor

7

Differential Revision: D50840897

gnahzg

Fix device propagation, tests for cpu sharding

33

Differential Revision: D51309697 Privacy Context Container: L1138451

IvanKobzarev

test use newest hypothesis

1

Differential Revision: D51604284

henrylhtsang

In pytorch embedding module, it has scale_grad_by_freq option, to scale gradients by the inverse of frequency of the words in the mini-batch. I am thinking if this is also doable...

fangleigit

StreamSyncTensor implicit pipelining for TorchRec [Prototype]

2

Provide implicit pipelining in TorchRec by using streams. StreamSyncTensor will continue on its current stream if all the incoming tensors are StreamSync tensors on that Stream. Otherwise it will sync...

YLGH

upgrade pyre version in `fbcode/torchrec` - batch 1

1

Differential Revision: D51504888

henrylhtsang

Remove call with block_bucketize_pos

3

Summary: Fix issue with https://fb.workplace.com/groups/1069285536500339/permalink/6853872844708217/ Differential Revision: D51479073

gnahzg

Initial PropagatingAsyncColletiveTensor

1

Summary: See https://docs.google.com/document/d/1hVIMaR6h8S7neMIeTQ3Y5pMKUTiU5IekEANklkABRnk/edit Reviewed By: ezyang Differential Revision: D51312623

YLGH