Deduplicate fp32 weights under torch autocast and ZeRO3 #7651
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
When torch autocast is enabled, model weights are already in fp32 and can be directly updated by the optimizer with fp32 gradients. It is a waste of accelerator memory to keep another copy, also in fp32, as the master weight. Use aliases to the so-called-"fp16" params as the master weights to save memory. It applies only when no optimizer offloading (either CPU or NVMe) or swapping mechanisms is enabled. Signed-off-by: Junjie Mao <[email protected]>
tohtana
approved these changes
Oct 27, 2025
Collaborator
tohtana
left a comment
tohtana
left a comment
There was a problem hiding this comment.
Thank you @eternalNight! I actually encountered this issue and was wondering how to fix it.
This is definitely a significant improvement.
4 tasks
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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Outdated suggestions cannot be applied.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.
When torch autocast is enabled, model weights are already in fp32 and can be directly updated by the optimizer with fp32 gradients. It is a waste of accelerator memory to keep another copy, also in fp32, as the master weight.
Use aliases to the so-called-"fp16" params as the master weights to save memory. It applies only when no optimizer offloading (either CPU or NVMe) or swapping mechanisms is enabled.
Using https://gist.github.com/eternalNight/3c2cf8c703f1e9e7742d3b7f9e1edae3 (which enables torch autocast) as an example, the memory profile of the training startup phase is as follows:
With this PR, the master weights no longer instantiate:
This is also true when DeepCompile is enabled:
When torch autocast is disabled, the master weights are preserved: