Skip to content

Use ViewKV flash attention for Qwen prefix prefill - #477

Merged
0xShug0 merged 1 commit into
mainfrom
fix/higgs-qwen-prefix-viewkv
Sep 7, 2026
Merged

Use ViewKV flash attention for Qwen prefix prefill#477
0xShug0 merged 1 commit into
mainfrom
fix/higgs-qwen-prefix-viewkv

Conversation

@0xShug0

@0xShug0 0xShug0 commented Sep 7, 2026

Copy link
Copy Markdown
Owner

Summary:

  • Route Qwen prefix-prefill flash attention through the ViewKV helper.
  • Keep the no-prefix FlashGrouped path separate and unchanged.
  • Avoid materializing grouped K/V for the prefix case.

Validation:

  • Built audiocpp_cli with the debug CUDA build.
  • A/B regression test on affected models

@0xShug0
0xShug0 merged commit c55826b into main Sep 7, 2026
6 checks passed
@0xShug0
0xShug0 deleted the fix/higgs-qwen-prefix-viewkv branch September 7, 2026 04:35
@0xShug0
0xShug0 restored the fix/higgs-qwen-prefix-viewkv branch September 7, 2026 04:53
@0xShug0
0xShug0 deleted the fix/higgs-qwen-prefix-viewkv branch September 7, 2026 04:53
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.

1 participant