Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Sdpa pcc improvement #13997

Merged
merged 2 commits into from
Oct 18, 2024
Merged

Sdpa pcc improvement #13997

merged 2 commits into from
Oct 18, 2024

Conversation

caixunshiren
Copy link
Contributor

@caixunshiren caixunshiren commented Oct 18, 2024

Ticket

#13364
#13866

Problem description

Greatly improved pcc of sdpa decode for large sequence length. PCC improved from 0.73 to 0.995+ for 128K sequence length.

What's changed

  • fixed fp32 accumulation error
  • added exp approx program option, which greatly improves pcc

Checklist

@caixunshiren caixunshiren merged commit bf48f8e into main Oct 18, 2024
7 checks passed
@caixunshiren caixunshiren deleted the sdpa-pcc-improvement branch October 18, 2024 18:53
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
Development

Successfully merging this pull request may close these issues.

2 participants