torchrec

torchrec copied to clipboard

fix eb _create_input_dist

1

Summary: keep it consistent with https://fburl.com/code/n6v2jb8f . We need to set these variables as empty to re-create_input_dist Differential Revision: D53167840

uciyc123

fix unsharded FeatureProcessedEmbeddingBagCollection when passing in excessive features

2

Summary: when extra features(keys) are passed to unsharded FeatureProcessedEmbeddingBagCollection, it needs to add dummy 1.0 weights to those excessive features to keep the kjt.weights.numel() same as kjt.values.numel() Reviewed By: YLGH...

bigning

sort the shard by perf instead of storage

2

Summary: We tried this idea before with default GreedyPartitioner, but we didn't land it, because: 1. we don't see significantly QPS improvment 2. Concern that if we don't place the...

bigning

suppress errors in `torchrec`

1

Differential Revision: D53111536

henrylhtsang

test update release build

1

Differential Revision: D53024751

henrylhtsang

test gpu ci for distributed

7

Differential Revision: D52841757

henrylhtsang

Cannot compute the gradients of EmbeddingBagCollection.

9

I try to test the demo scripts with the following command. ```torchx run -s local_cwd dist.ddp -j 1x1 --script test_installation.py``` I try to print the gradients of embedding_bag_collection. I can...

ZhuYuJin

CPU Unit Test CI to 15 minutes

1

Differential Revision: D52631694

PaulZhang12

MCH LRU Eviction Policy (#1598)

2

Summary: LRU eviction policy with user-variable decay exponent (e.g. decay_exponent=1 is LRU with linear distance). Reviewed By: dstaay-fb Differential Revision: D52100910

gsethi523

Update cost model to include prefetch_compute

1

Summary: Updated version of D50162035 that uses the new CacheStatistics interface, and updated for more sharding types. Jobs with large prefetch streams can be slowed down by large prefetch_compute blocking...

damianr99