Better solution to resolve DDP error when `tune-visual`

[youliangtan]

This is currently a hack to unblock things. Need better solution

Originally posted by @youliangtan in #257 (comment)

Need a more elegant solution to resolve:

making sure all `forward` function outputs participate in calculating loss. 
If you already have done the above, then the distributed data parallel module wasn't able to locate the output tensors in the return value of your module's `forward` function. Please include the loss function and the structure of the return value of `forward` of your module when reporting this issue (e.g. list, dict, iterable).
Parameter indices which did not receive grad for rank 0: 437 438 439 440 441 442 443 444 445 446 447
 In addition, you can set the environment variable TORCH_DISTRIBUTED_DEBUG to either INFO or DETAIL to print out information about which particular parameters did not receive gradient on this rank as part of this error

when tune-visual is enabled.

[xwjiang2010]

@youliangtan Can you provide more context regarding this issue?

[youliangtan]

@xwjiang2010 iikr multigpu ddp will fail without this

            for param in self.eagle_model.vision_model.parameters():
                if param.requires_grad:
                    dummy_term = dummy_term + 0.0 * param.sum()
            eagle_embeds = eagle_embeds + dummy_term

It works, but just if anyone can have a better solution. It just isnt elegant