torchrec
torchrec copied to clipboard
pytorch
Pytorch domain library for recommendation systems
Differential Revision: D38637338
Summary: For batched dense embedding we incorrectly provide a `named_buffers` implementation due to inheritance. For batched dense embeddings the weight values should be provided only through `named_parameters` rather than `named_buffers`...
Summary: Incorporate Torchrec's OSS planner into train_module. The OSS planner can be enabled by setting "`use_torchrec_oss_planner`" flag to true in SharderOptions (https://fburl.com/code/xjlbrscm). When the flag is true, `run_planner` calls `run_oss_planner`...
Summary: This change adds a new `PooledEmbeddingsReduceScatterV` which uses `_reduce_scatter_v` and `_all_gather_v` in c10d for performing row-wise and table-wise row-wise sharding. Differential Revision: D37735264
Summary: Update torchrec docs to include docker Differential Revision: D38466887
when I use horovod to run deepFM, it error File "/home/maer/zhipeng.li/project/torch_rec_demo/torchrec_ctr/run_horovod_deepfm.py", line 451, in main(args) File "/home/maer/zhipeng.li/project/torch_rec_demo/torchrec_ctr/run_horovod_deepfm.py", line 380, in main model = init_model(device=device) File "/home/maer/zhipeng.li/project/torch_rec_demo/torchrec_ctr/run_horovod_deepfm.py", line 187, in init_model...
Summary: The current EmbeddingBagCollection/FusedEmbeddingBagCollection are only usable through the DistributedModelParallel wrapper which override common torch.nn.module APIs (named_parameters/state_dict) etc. However, this makes these modules inflexible, and sometimes unusable without using DMP....
Summary: Since we no longer rely on DistributedModelParallel (for composability piece), we need an alternative way of getting the fused optimizer. get_fused_optimizer implements this, logically it's the same as the...
## Desired behavior We want to make TorchRec sharding composable w/ other sharding/parallelism techniques. This practically means that after applying TorchRec sharding model characteristics remain the same (e.g. state_dict() doesn’t...
Summary: pruning sharding options in the greedy proposer like we did in the grid search proposer to reduce the search space. -> by so, number of proposals decrease and thus...
