Skip to content

whisper: validate get_rows support for cpu extra buffer #3323

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

Merged
merged 1 commit into from
Jul 14, 2025

Conversation

chaxu01
Copy link
Contributor

@chaxu01 chaxu01 commented Jul 14, 2025

This patch enables KleidiAI to accelerate the Q4_0 matmul operation when it shares weight data with the get_rows operator.

@ggerganov ggerganov merged commit 032697b into ggml-org:master Jul 14, 2025
53 checks passed
@ggerganov
Copy link
Member

Could you share some sample numbers of the performance before and after this change?

@chaxu01
Copy link
Contributor Author

chaxu01 commented Jul 14, 2025

Here's benchmark Comparison (GET_ROWS vs. Baseline) for Pixel 8:

Metric GET_ROWS Baseline % Difference
Load time (ms) 82.16 61.14 25.58%
Encode time (ms) 856.86 841.23 1.82%
Decode time (ms) 401.13 423.74 -5.64%
Batch decode time (ms) 303.76 370.42 -21.94%
Prompt time (ms) 2838.28 3711.67 -30.77%
Total time (ms) 4401.69 5350.61 -21.56%

@chaxu01
Copy link
Contributor Author

chaxu01 commented Jul 14, 2025

This PR has a dependency on llama.cpp PR #14676, which introduces KleidiAI support for the get_rows operator.

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