bert: fix layer norm epsilon value #1946
Merged
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
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Outdated suggestions cannot be applied.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.
ref https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2/blob/7dbbc90392e2f80f3d3c277d6e90027e55de9125/config.json#L13
This is a quick-and-dirty fix since this code is going to be replaced anyway. It would be more correct to read layer_norm_eps when we convert to GGUF, and load that hyperparameter from the GGUF at inference time.
The difference between an epsilon of 1e-6 in LLaMA 1 and 1e-5 in LLaMA-2 created a significant difference in perplexity, so they implemented this parameter to ggml_norm and ggml_rms_norm soon after LLaMA-2 came out, and until the switch to GGUF they defaulted to 5e-6, which was a suitable middleground, and allowed the user to customize the parameter at inference time via a command-line option.
The difference between 1e-5 and 1e-12 is certainly more significant... if only we had benchmarks for this code.