Skip to content

aiter-flash-attn: rename public sink param to s_aux#992

Closed
Abdennacer-Badaoui wants to merge 1 commit into
huggingface:mainfrom
Abdennacer-Badaoui:aiter-flash-attn-rename-sink-to-s_aux
Closed

aiter-flash-attn: rename public sink param to s_aux#992
Abdennacer-Badaoui wants to merge 1 commit into
huggingface:mainfrom
Abdennacer-Badaoui:aiter-flash-attn-rename-sink-to-s_aux

Conversation

@Abdennacer-Badaoui

Copy link
Copy Markdown
Member

transformers' FA introspection only recognizes s_aux (vllm-fa3) or learnable_sink (FA4) as the sink-param name. Our public arg was sink, so transformers silently dropped the sink tensor on the way to the kernel and attention ran without sinks (visible as degenerate repetitive output on gpt-oss). Renaming the public param to s_aux (internal kernel path unchanged) fixes the wire-up.

vasqu
vasqu previously approved these changes Jun 23, 2026

@vasqu vasqu left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Good from my side not sure whether we raise versions here? cc @danieldk

Is versioning only related to layers or anything that faces the public API in some way?

@vasqu vasqu dismissed their stale review June 23, 2026 15:28

We will go with a different approach where we allow different namings for the same parameter

@Abdennacer-Badaoui

Abdennacer-Badaoui commented Jun 23, 2026

Copy link
Copy Markdown
Member Author

No need for this. We will allow multiple different names for the same argument here : huggingface/transformers#45153

@Abdennacer-Badaoui Abdennacer-Badaoui deleted the aiter-flash-attn-rename-sink-to-s_aux branch June 23, 2026 15:34
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants