flash_paged: s_aux may not exist#40434
Merged
ArthurZucker merged 1 commit intomainfrom Aug 26, 2025
Merged
Conversation
Some implementations (i.e., https://huggingface.co/kernels-community/vllm-flash-attn3) support an `s_aux` arg for attention sinks, but others (https://huggingface.co/kernels-community/flash-attn) do not. If s_aux is present in the kwargs, we forward it, otherwise we don't. The user will still get an error if they use a model like gpt-oss-20b with an implementation that does not support `s_aux`, but models that don't use it won't error out. For example, [this is currently failing](https://github.com/huggingface/transformers/blob/399cd5c04b11ba3f740b4f76e8067326786405cc/examples/pytorch/continuous_batching.py#L16) because we are sending `s_aux: None` in the dict.
|
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. |
ArthurZucker
approved these changes
Aug 26, 2025
Collaborator
ArthurZucker
left a comment
ArthurZucker
left a comment
There was a problem hiding this comment.
absolutely! thanks for the fix
4 tasks
ghost
mentioned this pull request
2 tasks
ArthurZucker
pushed a commit
that referenced
this pull request
Apr 23, 2026
…5589) Guard `s_aux` cast in `flash_attention_forward` for sink-less models `flash_attention_forward` unconditionally called `s_aux.to(query.dtype)`, which crashed with `AttributeError: 'NoneType' object has no attribute 'to'` for models that don't use attention sinks (e.g. Gemma). Mirrors the parallel guard added in #40434 for `flash_paged.py`. Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
ArthurZucker
pushed a commit
that referenced
this pull request
Apr 23, 2026
…5589) Guard `s_aux` cast in `flash_attention_forward` for sink-less models `flash_attention_forward` unconditionally called `s_aux.to(query.dtype)`, which crashed with `AttributeError: 'NoneType' object has no attribute 'to'` for models that don't use attention sinks (e.g. Gemma). Mirrors the parallel guard added in #40434 for `flash_paged.py`. Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
tarekziade
pushed a commit
that referenced
this pull request
Apr 23, 2026
…5589) Guard `s_aux` cast in `flash_attention_forward` for sink-less models `flash_attention_forward` unconditionally called `s_aux.to(query.dtype)`, which crashed with `AttributeError: 'NoneType' object has no attribute 'to'` for models that don't use attention sinks (e.g. Gemma). Mirrors the parallel guard added in #40434 for `flash_paged.py`. Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
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
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.
Some implementations (i.e.,
https://huggingface.co/kernels-community/vllm-flash-attn3) support an
s_auxarg for attention sinks, but others(https://huggingface.co/kernels-community/flash-attn) do not. If s_aux is present in the kwargs, we forward it, otherwise we don't.
The user will still get an error if they use a model like gpt-oss-20b with an implementation that does not support
s_aux, but models that don't use it won't error out. For example, this is currently failing because we are sendings_aux: Nonein the dict. We get: