Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Fix typo in chronicles-prequel.md #65

Open
wants to merge 1 commit into
base: master
Choose a base branch
from
Open

Fix typo in chronicles-prequel.md #65

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion train/tr11-176B-ml/chronicles-prequel.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -486,7 +486,7 @@ Also note that the model size isn't always the same as the number of layers had

* Sec/it throughput at iteration 2

As you can see the 80GB is totally unnecessary for MBS=1 as we are bound by compute of each gpu and we barely use half the gpu memory and trying to pack more on each gpu slows the ensemble down. This is of course thanks to ZeRO which shards all fp32 optim+grad+params over all gpus - so the more gpus you use the less memory is needed to accomodate the same model size, regardless of DP/TP/PP topology. (with MBS=1 that is so that the activations don't take too much memory)
As you can see the 80GB is totally unnecessary for MBS=1 as we are bound by compute of each gpu and we barely use half the gpu memory and trying to pack more on each gpu slows the ensemble down. This is of course thanks to ZeRO which shards all fp32 optim+grad+params over all gpus - so the more gpus you use the less memory is needed to accommodate the same model size, regardless of DP/TP/PP topology. (with MBS=1 that is so that the activations don't take too much memory)

This table doesn't take into account batch size rampup which needs to be divisible by DP as it progressed from 32, 64, ... so really we have an additional constraint of `DP % 4 = 0` and `GBS % 32 = 0`.

Expand Down