torchrec
torchrec copied to clipboard
pytorch
Pytorch domain library for recommendation systems
Reviewed By: inseokhwang Differential Revision: D54686975
Summary: * introduce static flags for weighted and cuda path based on tbe attributes instead of runtime value, more fx friendly and we will remove this wrap for tensor based...
Summary: ### Overview https://fburl.com/gdoc/efm9by7o - Generate table-wise sharding for LocationType.HBM and row-wise sharding for LocationType.DI_HOST - So we can directly edit the DIShardingPass with extra arguments - For table-wise sharding,...
Summary: to address: https://github.com/pytorch/torchrec/issues/1736 Differential Revision: D54547391
Differential Revision: D54490404
Summary: X-link: https://github.com/facebook/sapp/pull/108 X-link: https://github.com/pytorch/torchsnapshot/pull/168 X-link: https://github.com/pytorch/tnt/pull/717 X-link: https://github.com/facebook/pyre-check/pull/818 Reviewed By: connernilsen Differential Revision: D54436435
Summary: In the current implementation of `from_dense` method `self.lengths` always created on cpu. This is the problem, because methods that expects both `values` and `lengths` to be on the same...
Summary: As titled Reviewed By: jingsh Differential Revision: D54313506
Summary: Error is reported when using a combination of rowwise/tablewise sharding in SQEC. This is diff fix this error in SQEC. A refactor is needed and in the following diff...
Summary: torch.cat fails on empty list, guarding this case. Differential Revision: D54305327
