Paul Zhang

[email protected]

Results 8 comments of


                                            Paul Zhang

error with lengths = [1, 1, 1, 1,...., 1, 1, 1]

Can we see the whole stack trace please?

weight init in embedding_config should depend on embedding_dim

Hmm, fair point. Put out a PR?

gpu version `int_nbit_split_embedding_codegen_lookup_function` not support `indices` and `offsets` with uint64 dtype

It what cases would it be necessary to use uint64?

quantize_embeddings + KeyedJaggedTensor+ vbe cannot work

I don't believe VBE + quantized EBC is yet supported. Quantized EBC uses a completely different FBGEMM TBE than the standard EBC for training

Long running time when using torch.compile with DistributedModelParallel and dynamo errors

I don't believe torch.compile and TorchRec/DMP are fully compatible yet, @IvanKobzarev

`AttributeError: 'NoneType' object has no attribute '_dynamo_weak_dynamic_indices'` when using row-wise sharding

Dynamo isn't yet supported for every sharding combination in TorchRec, cc @IvanKobzarev to verify

Fix dynamic embedding bugs that may lead to segmentation faults

Please address CI failures and then you are good

Fix dynamic embedding bugs that may lead to segmentation faults

CI looks good, previously flaky test. Feel free to merge!