torchrec

torchrec copied to clipboard

fix model split boundary

2

Summary: As title Differential Revision: D56148751

PaulZhang12

prototype sharding by device group

1

Summary: This is a prototype that support tables with same sharding type but place on different device group. e.g some table-wise sharded table placed on CPU, some table-wise sharded table...

gnahzg

add tag for dataloader_iter in train_pipeline

2

Summary: # context In a recent experiemnt with HSNN, we moved the KJT concat to the dataloader. Adding this tag for better understanding the impact. Differential Revision: D56124045

TroyGarden

DMP doesn't broadcast DataParallel ShardingType embedding table from the process with rank 0 to all other processes

13

DMP should broadcast DataParallel ShardingType embedding table param from the process with rank 0 to all other processes in the group to make sure that all model replicas start from...

tiankongdeguiji

Wrap DP tables with DDP after reset parameters

3

Summary: # Problem Problem is once we wrap DP tables with DDP, the parameters are the same, but not synced. So if we reset the parameters in the table without...

henrylhtsang

batch_size * lengths for export

6

Differential Revision: D55602984

PaulZhang12

KJT dynamo path for conditions

4

Summary: The change is no-change for eager path. **Dynamo path** Dynamo tracing can not pass through various conditions => making dynamo path without those conditions. We skip asserts for dynamo...

IvanKobzarev

add bag reverse mapping for block_bucketize kernel

2

Summary: X-link: https://github.com/pytorch/FBGEMM/pull/2477 * we add the per indices bucketize bag reverse index return for inference batching needs. Differential Revision: D55492793

YazhiGao

Enable columnwise in sharders

2

Summary: As titled Reviewed By: henrylhtsang Differential Revision: D55381125

gnahzg

dynamo eager ShardedEBC

10

Differential Revision: D55080370

IvanKobzarev