qwen4exp : build the QSA selection without the kq_mask by sanmai · Pull Request #30965 · ggml-org/llama.cpp
retained text
gigabytes of VRAM saved (Qwen Flash Next) "This is one of the things Strata does to squeeze the most juice from the experts' cache."
document fetched 2026-10-24 10:43 · canonical https://reddit.com/r/LocalLLaMA/comments/1xdmek0/qwen4exp_build_the_qsa_selection_without_the_kq