Remove unnecessary fp32/bf16 conversion #1169
Merged
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
No
torch.distributed
operation that would require a bf16 -> fp32 conversion is performed in the_split
function, so it's safe to remove this.Since
fp32_allreduce=True
is necessary for combining zero 1 & bf16, this should result in less memory allocation for common zero 1 + MP + PP use cases (saw ~10% drop on 125M model, albeit with batch size 4)