torchrec

torchrec copied to clipboard

Added model config for supporting SparseNN, TowerSparseNN, TowerCollectionSparseNN models.

1

Summary: Added a model configuration that supports SparseNN, TowerSparseNN, TowerCollectionSparseNN models. Future commits will add support for DeepFM and DLRM models. Differential Revision: D76833673

SSYernar

Benchamrk framework for torchrec

6

Summary: Benchmark framework for MPZCH Rollback Plan: Differential Revision: D76150895

lizhouyu

Hey team, we have a working pipeline that uses torchrec 1.1.0. When we bump the version to 1.2.0, we are getting the following error (removed path prefixes): ``` File "sharded_sparse_and_dense_sequence_module.py",...

lneves-sc

Use Index Select instead of permute operation reduction in remote ro compute time

1

Summary: Did some offline benchmarking and found using index select is significantly faster than existing fbgemm.permute_embeddings when running KTRegroupAsDict for Ads model remote Ro. As a result pushing this up...

taylorKempWork

Added Optimizer configuration that supports optimizer type, learning rate, momentum, and weight decay configurations.

2

Summary: This commit introduces enhancements to the optimizer configuration in TorchRec. It now supports specifying the optimizer type, learning rate, momentum, and weight decay. These changes provide more flexibility and...

SSYernar

OSS TorchRec Internal MPZCH modules (#3017)

2

Summary: ### Major changes - Copy the following files from `fb` to corresponding location in the `torchrec` repository - `fb/distributed/hash_mc_embedding.py → torchrec/distributed/hash_mc_embedding.py` - `fb/modules/hash_mc_evictions.py → torchrec/modules/hash_mc_evictions.py` - `fb/modules/hash_mc_metrics.py → torchrec/modules/hash_mc_metrics.py`...

lizhouyu

add support for List[torch.Tensor]

1

Summary: list[torch.tensor] states were not handled correctly, this diff adds the requisite support TODO: in next diff we will optimize the collective calls such that we are not all gathering...

iamzainhuda

Test commit

7

Differential Revision: D76457454

aliafzal

Simplify fqns_to_feature_names in Delta Tracker

3

Summary: Each table in collection will have FQN. We can simplify the logic with this assumption to avoid two iterations. Existing tests passed with this logic. Differential Revision: D76432354

kausv

I want to use CPU-based distributed approach to train a small recommendation model. Is there a demo available for me to refer to?

1

Lenan22