Grad Norm Differences Across Nodes

[EugenHotaj]

Continuing the discussion from #2172 (thanks @mirceamironenco, @ebsmothers for the fix!).

We have a run on the exact same dataset / hparams except we change the number of nodes from 8->2->1. We noticed that when we reduce the number of nodes the gradient norm goes up:

Here is an 8 node run:

Here is a 2 node run:

Here is a 1 node run:

We can see the grad norm at initialization is ~4x different between 8 node and 1 node run. With the fix in #2172, I would expect the grad norms to be similar regardless of the world size. The only difference between the runs is the global batch size (64 on 1 node, 512 on 8 nodes), but I would not expect this to cause such a big difference.

Is it possible there are still some issues in how we compute / scale the gradients?

[EugenHotaj]

I think this may just be a logging issue actually. To get the correct grad_norm we need to call grad_norm.full_tensor() like they do in torchtitan here: https://github.com/pytorch/torchtitan/blob/90567fc9827ffdf17bdd0349cd5276c662d0769a/torchtitan/utils.py#L396

I think right now we're just reporting the (unreduced) total norm from rank 0. If I call full_tensor() in our local runs, I get the same norm regardless of the world size + matches what we're seeing in an identical run in NeMo.

[ebsmothers]

Thanks @EugenHotaj for the issue and the update. I agree that full_tensor() should give the correct grad. Just ran a quick test to make sure there's no major perf hit to doing this and it looks good. Do you wanna open a PR with the fix? (If not let me know and I am happy to do so)

[EugenHotaj]

@ebsmothers sorry just saw your comment / pull request. Thanks for the quick fix!

[ebsmothers]

Closing now that #2248 has landed