log learning rate #937
Closed
log learning rate #937
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
fegin
reviewed
Mar 6, 2025
Co-authored-by: Chien-Chin Huang <[email protected]>
tianyu-l
reviewed
Mar 6, 2025
Contributor
tianyu-l
left a comment
tianyu-l
left a comment
There was a problem hiding this comment.
sounds good to me.
Could you please include a screenshot of the results? Either TB or WandB is fine.
Co-authored-by: tianyu-l <[email protected]>
Contributor
Author
Here's a screenshot on WandB (before the log's name change).
Currently that PR does not includes changes to the Exp Tracker, only to the logger. So it's orthogonal to this PR I think. |
Contributor
|
sorry what is "Exp Tracker"? |
Contributor
Author
|
Ah yes sorry I was confused and missed that part. Those are the same thing. I will close this PR in favor of that PR. |
tianyu-l
pushed a commit
that referenced
this pull request
Jul 31, 2025
This PR adds learning rate logging. There was a previous attempt to implement this in an [earlier PR](#937), but that one was ultimately **closed**. This version ensures that LR logging works properly, I verified it using the WSD scheduler that was recently added in [another PR](#938). <img width="1842" height="730" alt="image" src="https://github.com/user-attachments/assets/8f23674a-d689-4cc2-9d9b-30bff4e63f3b" /> One design consideration here is that torchtitan supports multiple optimizers and learning rate schedules, each potentially having its own LR. However, in practice, I believe that 99.9999% of use cases will use a single LR. Given that, the logging works as follows: - If there is only one learning rate, it gets logged directly under the main charts as `lr`. - If there are multiple learning rates, they are logged under a separate section, each with its corresponding label. Alternatively, we could have ignored the multi-LR case and always logged a single LR, but I prefer this approach since it handles both scenarios robustly with minimal extra code. Happy to adjust if others have a strong preference for simplicity over robustness.
bentherien
pushed a commit
to bentherien/torchtitan_
that referenced
this pull request
Aug 5, 2025
This PR adds learning rate logging. There was a previous attempt to implement this in an [earlier PR](pytorch#937), but that one was ultimately **closed**. This version ensures that LR logging works properly, I verified it using the WSD scheduler that was recently added in [another PR](pytorch#938). <img width="1842" height="730" alt="image" src="https://github.com/user-attachments/assets/8f23674a-d689-4cc2-9d9b-30bff4e63f3b" /> One design consideration here is that torchtitan supports multiple optimizers and learning rate schedules, each potentially having its own LR. However, in practice, I believe that 99.9999% of use cases will use a single LR. Given that, the logging works as follows: - If there is only one learning rate, it gets logged directly under the main charts as `lr`. - If there are multiple learning rates, they are logged under a separate section, each with its corresponding label. Alternatively, we could have ignored the multi-LR case and always logged a single LR, but I prefer this approach since it handles both scenarios robustly with minimal extra code. Happy to adjust if others have a strong preference for simplicity over robustness.
joellidin
pushed a commit
to one-covenant/torchtitan
that referenced
this pull request
Aug 8, 2025
This PR adds learning rate logging. There was a previous attempt to implement this in an [earlier PR](pytorch#937), but that one was ultimately **closed**. This version ensures that LR logging works properly, I verified it using the WSD scheduler that was recently added in [another PR](pytorch#938). <img width="1842" height="730" alt="image" src="https://github.com/user-attachments/assets/8f23674a-d689-4cc2-9d9b-30bff4e63f3b" /> One design consideration here is that torchtitan supports multiple optimizers and learning rate schedules, each potentially having its own LR. However, in practice, I believe that 99.9999% of use cases will use a single LR. Given that, the logging works as follows: - If there is only one learning rate, it gets logged directly under the main charts as `lr`. - If there are multiple learning rates, they are logged under a separate section, each with its corresponding label. Alternatively, we could have ignored the multi-LR case and always logged a single LR, but I prefer this approach since it handles both scenarios robustly with minimal extra code. Happy to adjust if others have a strong preference for simplicity over robustness.
joellidin
pushed a commit
to one-covenant/torchtitan
that referenced
this pull request
Aug 8, 2025
This PR adds learning rate logging. There was a previous attempt to implement this in an [earlier PR](pytorch#937), but that one was ultimately **closed**. This version ensures that LR logging works properly, I verified it using the WSD scheduler that was recently added in [another PR](pytorch#938). <img width="1842" height="730" alt="image" src="https://github.com/user-attachments/assets/8f23674a-d689-4cc2-9d9b-30bff4e63f3b" /> One design consideration here is that torchtitan supports multiple optimizers and learning rate schedules, each potentially having its own LR. However, in practice, I believe that 99.9999% of use cases will use a single LR. Given that, the logging works as follows: - If there is only one learning rate, it gets logged directly under the main charts as `lr`. - If there are multiple learning rates, they are logged under a separate section, each with its corresponding label. Alternatively, we could have ignored the multi-LR case and always logged a single LR, but I prefer this approach since it handles both scenarios robustly with minimal extra code. Happy to adjust if others have a strong preference for simplicity over robustness.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 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.
Simply log all the learning rates for all parameter groups of all schedulers.