Introduce the scale enum flag in Embedding layer for LLM embedding. #909
Conversation
This file contains 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
|
@ruomingp Could you review? From 970 |
ruomingp
reviewed
Jan 8, 2025
axlearn/common/layers.py
Outdated
| https://github.com/google-deepmind/gemma/blob/0d6ae857591248422127ca14c027909546362e6a/gemma/modules.py#L80 | ||
| """ | ||
|
|
||
| UNIT = 1 |
There was a problem hiding this comment.
Nit: use str values to be more readable.
Suggested change
| UNIT = 1 | |
| UNIT = "unit" |
The activation component should roughly have a magnitude of 1. Since the embedding tensor is initialized with a scale of `1/sqrt(dim)`, the activation is multiplied by `sqrt(dim)` to maintain the desired scale. e.g. Gemma [1] [1] https://github.com/google-deepmind/gemma/blob/0d6ae857591248422127ca14c027909546362e6a/gemma/modules.py#L80 In addition, unsloth [2] discovered that `sqrt(dim)` needs to be computed in float32. [2] Sec 3 in https://unsloth.ai/blog/gemma-bugs TODO(axlearn-team): Use UNIT scale enum for AFM+. This will require re-sweeping hyperparameters (e.g., learning rate).
ruomingp
approved these changes
Jan 8, 2025
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Outdated suggestions cannot be applied.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.
The activation component should roughly have a magnitude of 1. Since the embedding tensor is initialized with a scale of
1/sqrt(dim), the activation is multiplied bysqrt(dim)to maintain the desired scale. e.g. Gemma [1][1] https://github.com/google-deepmind/gemma/blob/0d6ae857591248422127ca14c027909546362e6a/gemma/modules.py#L80
In addition, unsloth [2] discovered that
sqrt(dim)needs to be computed in float32. [2] Sec 3 in https://unsloth.ai/blog/gemma-bugsTODO(axlearn-team): Use UNIT scale enum for AFM+. This will require re-sweeping hyperparameters (e.g., learning rate).