torchrec

torchrec copied to clipboard

RuntimeError: table-wise sharding on a single EmbeddingBag is not supported yet

3

I saw this error message when I was trying the `EmbeddingBagSharder`. I read the code and found table-wise sharding is not supported in `ShardedEmbeddingBag`. I believe this is a bug...

jiannanWang

Copy sparse, dense and labels from the right source when shuffling Criteo data

1

When dense and spare features are located in different directories, i.e., `input_dir_labels_and_dense != input_dir_sparse` as the input to `shuffle()` function, this method throws an error in [L642](https://github.com/pytorch/torchrec/blob/71d51c8764c141ff8d849f73bcf06548ffad36c4/torchrec/datasets/criteo.py#L642) like: ``` FileNotFoundError:...

janekl

Allow DMP to pass prefix fused params to sharded modules

4

Summary: As title Differential Revision: D38881744

xing-liu

fix tower input dist

2

Summary: before this change, for the shardedEmbeddingTower, in the input_dist, we only do the cross_pg dist. Then in the compute() step, we call the tower.embedding() to do the intra_pg input_dist...

bigning

tolist() support for FunctionalTensor

1

Summary: Support tolist() for FunctionalTensor for KJT in torch.export Reviewed By: ezyang Differential Revision: D53731064

PaulZhang12

Update PositionWeightedModule to make it jit trace compitable

1

Summary: In ``` torch.ops.fbgemm.offsets_range(features[key].offsets().long(), torch.numel(features[key].values()) ``` This part ``` torch.numel(features[key].values() ``` will be traced into constant Reviewed By: snabelkabiya Differential Revision: D53744703

ZhengkaiZ

Back out "Small syntactic changes for dynamo compatibility"

2

Summary: Original commit changeset: 013ae157559e Original Phabricator Diff: D53854369 See Details in https://fb.workplace.com/groups/mbf.users/permalink/1829315864212395/ This is regressing qps in model Reviewed By: zzzwen, henrylhtsang Differential Revision: D54138464

henrylhtsang

Introduce pt2 ir export unit test

2

Summary: Introduce unit testing torch.export Differential Revision: D53426083

PaulZhang12

Remove `device` in recat torch.Tensor

1

Summary: As titled Reviewed By: IvanKobzarev Differential Revision: D54119559

MatthewWEdwards

Improve perf of _block_bucketize_sparse_features_cuda_kernel2 - for pooled emb

1

Summary: X-link: https://github.com/pytorch/FBGEMM/pull/2346 This diffs tries to parallelize _block_bucketize_sparse_features_cuda_kernel2 more (the kernel that bucketize the data vector) based on the assumption that the order of id in an id list...

AlbertDachiChen