refactor: remove gate normalization in DecoderUnit - #222
Merged
Conversation
Simplify the gate computation by removing the normalization step. The gate values are now used directly from softmax scatter without the additional sum normalization.
There was a problem hiding this comment.
Pull request overview
This PR updates the DeepSeekMoE-style routed feed-forward gating inside DecoderUnit by removing the post-topk renormalization step, so routed expert weights are taken directly from the original softmax probabilities at the selected indices.
Changes:
- Removed
gate_primeconstruction and thegate = gate_prime / gate_prime.sum(...)normalization. - Use a single
gatetensor built viascatter_of the softmax probabilities at top-k indices.
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
| # gate_prime, gate: batch * site * routed | ||
| gate_prime = torch.zeros_like(similarity).scatter_(-1, top_k_indices, similarity.gather(-1, top_k_indices)) | ||
| gate = gate_prime / gate_prime.sum(dim=-1).unsqueeze(-1) | ||
| # gate, gate: batch * site * routed |
There was a problem hiding this comment.
The inline comment is now inaccurate/typoed after removing gate_prime (it reads gate, gate: ...). Please update it to describe the single gate tensor to avoid confusion when reading the routing logic.
Suggested change
| # gate, gate: batch * site * routed | |
| # gate: batch * site * routed (top-k routed gating weights) |
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Simplify the gate computation by removing the normalization step. The gate values are now used directly from softmax scatter without the additional sum normalization.