torchrec

torchrec copied to clipboard

Remove DMP state dict override and use post hook

1

Summary: Instead of overriding nn.Module's state dict implementation, satisfy DMP state dict needs by stripping DMP/DDP prefixes through a post hook. Differential Revision: D35826663

joshuadeng

Properly set up GPU/CUDA environment for CI

3

Summary: ATT Pytorch distributed when writing some GPU unit tests noticed that they didn't have access to GPUs. After some investigation they found that these additional lines were included in...

YLGH

create predictor gen code (py + c++)

1

Summary: in grpc's examples they include the gen code in case people want to run it without running the protoc commands Differential Revision: D35746611

s4ayub

Fix topology not properly set when running dlrm_main multinode

1

Luo-Liang

dlrm fixes

9

Add useful error check. Fix model training mode state so it's on when it should be. Stop adding validation samples to training set. Fix validation from running after a single...

samiwilf

Benchmark batching funcs

1

Summary: Benchmarks for: - combineSparse - combineDense (iobuf, ivalue) - combineEmbedding configurable batch size, number of features and number of requests for both Differential Revision: D35227798

s4ayub

Sequence embedding TWRW sharding

2

Summary: title Differential Revision: D34917433

bigning

request on behalf of @adamlerer (adding here to be able to reference) 1. the readme contains no information about how to actually use torchrec, or a problem description of what...

colin2328

Replace IValue::toString()->string() with IValue::toStringRef()

1

Summary: Same as previous diff just for pytorch/torchrec Reviewed By: jaybean-dev Differential Revision: D39707241

alexbeloi

Unified Authoring of Overlapped Optimizers, and per parameter optimizer settings

4

### Context and Problem: Optimizer overlap/fusion is a technique where upon calling backwards(), the optimizer step upon each parameter is applied as soon as its gradient is calculated. The benefits...

YLGH