Questions about DistTensor push local node features

Hi, thank you very much for your time to read my questions.

According to the answer in this postHow the distributed tensors work?, client access the local data via shared memory. Can I assume that I can update the DistTensor by updating the data shared memory without pushing? For example, in this test caseassign new value to DistTensor, after assigning new value “new_feat” to DistTensor “test1”, if I change the value in new_feat, the corresponding value in “test1” should also be changed, is that correct? So after changing the value of new_feat, the value in “test1” should also be updated, do I still need to update DistTensor using kvstore.push? Or if I do not push my update, remote server will be not able to pull the updated data?

I think g.ndata["test1"][nids] = new_feats is deep copy instead of reference. So update on new_feats will not be updated on DistTensor.

No, when DistTensor = Tensor is used, the data in the Tensor will be deep copied to the memory of DistTensor, which is shared among clients and servers on the same machine.

This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.