Fix rotary positional embedding dimension mismatch in Wan and SkyReels V2 transformers #12594
+11
−10
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
…s V2 transformers - Store t_dim, h_dim, w_dim as instance variables in WanRotaryPosEmbed and SkyReelsV2RotaryPosEmbed __init__ - Use stored dimensions in forward() instead of recalculating with different formula - Fixes inconsistency between init (using // 6) and forward (using // 3) - Ensures split_sizes matches the dimensions used to create rotary embeddings
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 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Outdated suggestions cannot be applied.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.
Fixes #12538
The original implementation had inconsistent formulas between init and forward:
In init: Dimensions calculated as h_dim = w_dim = 2 * (attention_head_dim // 6)
In forward: Split sizes calculated as attention_head_dim // 3
While mathematically similar, these can produce different results due to integer division, leading to potential tensor dimension mismatches when splitting the rotary embedding buffers.
Solution
Store the computed dimensions (t_dim, h_dim, w_dim) as instance variables in init and reuse them in forward().
Before submitting
documentation guidelines, and
here are tips on formatting docstrings.
Who can review?
@Warvito, @yiyixuxu
Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.