← sources

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

source: reddit (reddit:LocalLLaMA)published 2026-09-22 17:39ingested 2026-09-23 02:43

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