torchrec
torchrec copied to clipboard
pytorch
Pytorch domain library for recommendation systems
It seems torchrec does not support the combination of data parallelism and row-wise parallelism for embedding. I want to know is there a plan on it? Or is row-wise parallelism...
# Describe the bug A PSCollection should contain optimizer states besides weights. The optimizer states tensors are obtained directly [from EmbeddingCollection Module](https://github.com/pytorch/torchrec/blob/main/contrib/dynamic_embedding/src/torchrec_dynamic_embedding/ps.py#L165-L168). However, the [sharded_module.fused_optimizer.state_dict()['state']](https://github.com/pytorch/torchrec/blob/main/contrib/dynamic_embedding/src/torchrec_dynamic_embedding/ps.py#L153C32-L153C75) does not contain key `{table_name}.momentum2`...
Summary: As titled Differential Revision: D58834418
Summary: # context * convert `FeatureProcessedEmbeddingBagCollection` to custom op in IR export * add serialization and deserialization function for FPEBC * add an API for the `FeatureProcessorInterface` to export necessary...
Differential Revision: D59019375
Summary: Fusing states for each metric to reduce computation overhead. this will help **every** model that uses RecMetrics. By fusing state we no longer all gather per state, we see...
Summary: packaging issue importing `_permute_tensor_by_segments` - jagged_tensor.py refactor once jagged_tensor has the changes Differential Revision: D59014188
Summary: Implementing feature grouping in APS model generator to enable unified embedding onbaording to FM This diff introduced a new config section "unified_embedding" inside `AdsFeatureArchEntityConfig` and `AdsFeatureArchConfig` (for v1/v2 compatibility)...
Use `ast.unparse(ast.parse(code))` to normalize the source code. This ignores formatting differences and allows us to enable code formattor on `torch` upstream. As per: - https://github.com/pytorch/pytorch/pull/128594#issuecomment-2181558483 - https://github.com/pytorch/pytorch/pull/128594#discussion_r1649795743
Summary: # specs * inputs: a list of KTs a. the N-th KT has a shape of (batch_size, dimN), batch_size should be identical b. the N-th KT contains a list...