Fix Fast Gradient Clipping bias gradient calculation for three dim data #751
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
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D70823094 |
0189db5 to
d0a98cb
Compare
EnayatUllah
added a commit
to EnayatUllah/opacus
that referenced
this pull request
Apr 9, 2025
…ta (meta-pytorch#751) Summary: The bias grad calculation for three dim data was incorect. Let `G = g^Tg`, where `g`, of dimensions `Txd` be the per-sample activation gradient, where `T` is the number of tokens and `d` dimension. The per-sample gradient norm with respect to bias is `vec(G)^T vec(1)`, instead of the erroneous,`vec(G)^T vec(G)` before. This diff fixes it. Reviewed By: HuanyuZhang Differential Revision: D70823094
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D70823094 |
EnayatUllah
added a commit
to EnayatUllah/opacus
that referenced
this pull request
Apr 9, 2025
…ta (meta-pytorch#751) Summary: The bias grad calculation for three dim data was incorect. Let `G = g^Tg`, where `g`, of dimensions `Txd` be the per-sample activation gradient, where `T` is the number of tokens and `d` dimension. The per-sample gradient norm with respect to bias is `vec(G)^T vec(1)`, instead of the erroneous,`vec(G)^T vec(G)` before. This diff fixes it. Reviewed By: HuanyuZhang Differential Revision: D70823094
d0a98cb to
8f993b6
Compare
…ta (meta-pytorch#751) Summary: Pull Request resolved: meta-pytorch#751 The bias grad calculation for three dim data was incorect. Let `G = g^Tg`, where `g`, of dimensions `Txd` be the per-sample activation gradient, where `T` is the number of tokens and `d` dimension. The per-sample gradient norm with respect to bias is `vec(G)^T vec(1)`, instead of the erroneous,`vec(G)^T vec(G)` before. This diff fixes it. Reviewed By: HuanyuZhang Differential Revision: D70823094
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D70823094 |
8f993b6 to
d2d6e41
Compare
EnayatUllah
added a commit
to EnayatUllah/opacus
that referenced
this pull request
Apr 9, 2025
…ta (meta-pytorch#751) Summary: The bias grad calculation for three dim data was incorect. Let `G = g^Tg`, where `g`, of dimensions `Txd` be the per-sample activation gradient, where `T` is the number of tokens and `d` dimension. The per-sample gradient norm with respect to bias is `vec(G)^T vec(1)`, instead of the erroneous,`vec(G)^T vec(G)` before. This diff fixes it. Reviewed By: HuanyuZhang Differential Revision: D70823094
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D70823094 |
Contributor
|
This pull request has been merged in 7264cd7. |
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.
Summary:
The bias grad calculation for three dim data was incorect.
Let
G = g^Tg, whereg, of dimensionsTxdbe the per-sample activation gradient, whereTis the number of tokens andddimension.The per-sample gradient norm with respect to bias is
vec(G)^T vec(1), instead of the erroneous,vec(G)^T vec(G)before. This diff fixes it.Reviewed By: HuanyuZhang
Differential Revision: D70823094