# Performance tune for gemma4-26b-a4b flash attention shape. by frobnitzem · Pull Request #28450 · ggml-org/llama.cpp · GitHub — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/d01dc9ccfdff42dd

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-09-22 17:39 ingested 2026-09-23 02:43

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1wng98w/performance_tune_for_gemma426ba4b_flash_attention/)

## retained text



```


Gemma 26B A4B speedup
```



document fetched 2026-09-23 00:44 · canonical https://reddit.com/r/LocalLLaMA/comments/1wng98w/performance_tune_for_gemma426ba4b_flash_attention
