[Known Issue] Deepspeed multiple loss support

[ncilfone]

If running with multiple losses Deepseep will currently fail as retain_graph is not passed to the backward call within the Deepspeed engine thus preventing multiple backward calls within stoke.

Please use a single loss function for the time being until this can be patched -- in most simple situations with multiple losses simply add them...

PR is currently opened to fix this: microsoft/DeepSpeed#1149

[sualehasif]

We have been working with deepspeed and really need the retain_graph support. (We can't simply add our losses.) Waiting for this to be fixed :)

[ncilfone]

@sualehasif

Heads up, I think you might have wanted to respond to the PR on the deepspeed repo instead of this one!