Fix "inf" norm gradient clipping for FSDP2 #3199
Open
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
facebook-github-bot
added
the
CLA Signed
|
This pull request was exported from Phabricator. Differential Revision: D78326114 |
|
This pull request was exported from Phabricator. Differential Revision: D78326114 |
wz337
added a commit
to wz337/torchrec
that referenced
this pull request
Jul 16, 2025
Summary: "inf" norm calculation for FSDP2 is incorrect, as the `total_grad_norm` would always be 1.0 regardless of the gradients. This is due to `_batch_cal_norm` would calculate `total_grad_norm = total_grad_norm ** (1.0 / norm_type)`, even if the `norm_type` is `inf` norm. For `inf` norm, since `norm_type` is `torch.inf`, `total_grad_norm` would become `total_grad_norm ** (1.0 / torch.inf) `, which would always just be 1.0. Before the fix, line 220 is comparing `self._norm_type != torch.inf`, as `self._norm_type` is either float number of `"inf"`, it would always enter this if statement, which is incorrect for the infinity norm case. This issue is found during Shampoo War Room debugging for HSDP2 (context: https://fb.workplace.com/groups/3095840833991792/permalink/4084834545092411/). During ablation study, we found that DDP Fused Adam is on-par with HSDP2 Fused Adam when 2-norm is used for clipping (Figure 1), while HSDP2 Fused Adam behaves strangely (high variance between runs) when infinity norm is used for clipping (Figure 2). Figure 1 {F1980311095} Figure 2 {F1980311115} Differential Revision: D78326114
|
This pull request was exported from Phabricator. Differential Revision: D78326114 |
wz337
added a commit
to wz337/torchrec
that referenced
this pull request
Jul 16, 2025
Summary: Pull Request resolved: pytorch#3199 "inf" norm calculation for FSDP2 is incorrect, as the `total_grad_norm` would always be 1.0 regardless of the gradients. This is due to `_batch_cal_norm` would calculate `total_grad_norm = total_grad_norm ** (1.0 / norm_type)`, even if the `norm_type` is `inf` norm. For `inf` norm, since `norm_type` is `torch.inf`, `total_grad_norm` would become `total_grad_norm ** (1.0 / torch.inf) `, which would always just be 1.0. Before the fix, line 220 is comparing `self._norm_type != torch.inf`, as `self._norm_type` is either float number of `"inf"`, it would always enter this if statement, which is incorrect for the infinity norm case. This issue is found during Shampoo War Room debugging for HSDP2 (context: https://fb.workplace.com/groups/3095840833991792/permalink/4084834545092411/). During ablation study, we found that DDP Fused Adam is on-par with HSDP2 Fused Adam when 2-norm is used for clipping (Figure 1), while HSDP2 Fused Adam behaves strangely (high variance between runs) when infinity norm is used for clipping (Figure 2). Figure 1 {F1980311095} Figure 2 {F1980311115} Differential Revision: D78326114
Summary: "inf" norm calculation for FSDP2 is incorrect, as the `total_grad_norm` would always be 1.0 regardless of the gradients. This is due to `_batch_cal_norm` would calculate `total_grad_norm = total_grad_norm ** (1.0 / norm_type)`, even if the `norm_type` is `inf` norm. For `inf` norm, since `norm_type` is `torch.inf`, `total_grad_norm` would become `total_grad_norm ** (1.0 / torch.inf) `, which would always just be 1.0. Before the fix, line 220 is comparing `self._norm_type != torch.inf`, as `self._norm_type` is either float number of `"inf"`, it would always enter this if statement, which is incorrect for the infinity norm case. This issue is found during Shampoo War Room debugging for HSDP2 (context: https://fb.workplace.com/groups/3095840833991792/permalink/4084834545092411/). During ablation study, we found that DDP Fused Adam is on-par with HSDP2 Fused Adam when 2-norm is used for clipping (Figure 1), while HSDP2 Fused Adam behaves strangely (high variance between runs) when infinity norm is used for clipping (Figure 2). Figure 1 {F1980311095} Figure 2 {F1980311115} TODO: As discussed, we should be able to significantly simplify things by leveraging a similar implementation as what's used in TorchTitan (https://github.com/pytorch/torchtitan/blob/main/torchtitan/distributed/utils.py) and abstract out the gradient norm computation. Reviewed By: yoyoyocmu Differential Revision: D78326114
|
This pull request was exported from Phabricator. Differential Revision: D78326114 |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
Summary:
"inf" norm calculation for FSDP2 is incorrect, as the
total_grad_normwould always be 1.0 regardless of the gradients. This is due to_batch_cal_normwould calculatetotal_grad_norm = total_grad_norm ** (1.0 / norm_type), even if thenorm_typeisinfnorm. Forinfnorm, sincenorm_typeistorch.inf,total_grad_normwould becometotal_grad_norm ** (1.0 / torch.inf), which would always just be 1.0.Before the fix, line 220 is comparing
self._norm_type != torch.inf, asself._norm_typeis either float number of"inf", it would always enter this if statement, which is incorrect for the infinity norm case.This issue is found during Shampoo War Room debugging for HSDP2 (context: https://fb.workplace.com/groups/3095840833991792/permalink/4084834545092411/).
During ablation study, we found that DDP Fused Adam is on-par with HSDP2 Fused Adam when 2-norm is used for clipping (Figure 1), while HSDP2 Fused Adam behaves strangely (high variance between runs) when infinity norm is used for clipping (Figure 2).
Figure 1
{F1980311095}
Figure 2
{F1980311115}
Differential Revision: D78326114