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