Several fixes related to rotary position embeddings #35376
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
4 tasks
mseeger
force-pushed
the
fixrope_part1
branch
3 times, most recently
from
987d726 to
cdb18d5
Compare
First part of resolution of huggingface#35233 - Changes related to `position_embeddings` being a mandatory argument - Remove `position_ids` argument of `apply_rotary_pos_emb` - Replace `torch.stack` by `torch.cat`, former requires equal shapes - `esm`: RoPE depends on `position_ids`, which was ignored. - `gpt_neox`: Selection of attention compute type via class removed - `gptj`, `codegen`: RoPE must be applied per head, and some shape issues. - `nemotron`: `config.partial_rotary_factor` was not implemented.
mseeger
force-pushed
the
fixrope_part1
branch
from
cdb18d5 to
12a1a35
Compare
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Outdated suggestions cannot be applied.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.
What does this PR do?
position_embeddingsbeing a mandatory argumentposition_idsargument ofapply_rotary_pos_embtorch.stackbytorch.cat, former requires equal shapesesm: RoPE depends onposition_ids, which was ignored.Fix changes behavior, but should improve.
gpt_neox: Selection of attention compute type via class removedgptj,codegen: RoPE must be applied per head, and some shape issues.Probably changes behavior.
nemotron:config.partial_rotary_factorwas not implemented. This iswhy default changed to 1, so that behavior in default case does not change.
Fixes #35233. This is the first of two PRs providing the fix. I split it for easier reviewing.
Before submitting
Pull Request section?
to it if that's the case.
Shape mismatch in RoPE embeddings gpt_neox model when rotary_ndims is odd #35233
documentation guidelines, and
here are tips on formatting docstrings.
The second PR fixes the bug the issue is about, and I am adding new tests to confirm it is fixed.
Who can review?
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.