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Outdated suggestions cannot be applied.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.
This PR is an attempt at replacing the regular softmax for MoE gating with a Gumbel-Softmax function instead.
The paper Approximating Two-Layer Feedforward Networks for Efficient Transformers suggests that softmax may not be ideal for gating mechanism, arguing that it's too competitive and may result in some experts being disproportionately used in the selection process.
This PR is an experiment to see if replacing Softmax with Gumbel-Softmax can help with diversifying the expert selection process. By adding noise to the logits, the selection process will hopefully introduce enough fluctuation to the softmax probabilities that if one expert has a slightly higher logit value compared to another one, it won't be chosen every time.
I've added a new parameter,
temperature
, to control the amount of randomness added. This will need to be inserted in the model's config.json.