UPSTREAM PR #19095: tests : add GQA=20 FA test

[loci-dev]

Mirrored from ggml-org/llama.cpp#19095

Might be a good idea to have a test that exercises GQA=20 in order to catch any potential regressions.

[loci-review]

No summary available at this time. Visit Version Insights to review detailed analysis.

[loci-review]

Performance Review Report

Summary

No functions were identified for performance analysis between the base and target versions. This indicates that no meaningful performance changes occurred in this code revision.

Analysis

The absence of functions with significant response time or throughput time changes suggests that:

1. No Performance-Critical Modifications: The changes between versions did not impact the execution characteristics of performance-sensitive functions in the llama.cpp inference pipeline, including core operations like llama_decode(), matrix multiplication kernels, attention mechanisms, or KV cache management.

2. Stable Performance Profile: The computational characteristics of the codebase remain consistent across both versions, with no detectable differences in execution time for any functions.

3. Non-Performance Changes: Any modifications made between versions likely involved non-performance aspects such as code refactoring, documentation updates, bug fixes that don't affect execution paths, or changes to non-critical utility functions.

Conclusion

The target version maintains performance parity with the base version. No performance regressions or improvements were detected in any measured functions.

See the complete breakdown in Version Insights
Have questions? Tag @loci-dev to ask about this PR.