DGLHeteroGraph.multi_update_all with inplace=True

Hi, I am trying to train on a giant heterograph with multiple node & edge types. To deal with the memory issues; initially I tried message-passing only on a subgraph, yet the memory requirement keeps on increasing with every epoch.

I wanted to try DGLHeteroGraph.multi_update_all with inplace=True but I am getting the following error:
G.multi_update_all(funcs, ‘sum’, inplace=True)
TypeError: multi_update_all() got an unexpected keyword argument ‘inplace’

Without the argument ‘inpIace’ everything runs fine, but only for 3-4 epochs and then it runs out of memory.
I think inplace=True is not supported yet. Any suggestions?

For inplace=True you are right. It is not supported yet. Also, even if supported, “inplace=True” during training would still create a new tensor covering all nodes. The reason is to enable backpropagation. So it does not conserve memory as you would thought.

The constantly increasing memory consumption sounds strange as it sounds like a memory leakage. Would you mind sharing the minimal case that could reproduce this behavior?

Thanks for replying.
Turns out that my DGL version was not up to date and there was a recent(15-20 days back) update on github which fixed the memory leak issue. That is, memory requirement of my code is constant with every epoch, now.
I still need graph store support for single giant heterogeneous graph though.