qwen4exp: gather-based sparse attention for QSA decode (+ cuda top-k batching fix) - #165

qwen4exp: gather-based sparse attention for QSA decode (+ cuda top-k batching fix)#165

abdel-darwish-27 wants to merge 4 commits into

unslothai:qwen4exp/qwen3.8-flash-nextunslothai/llama.cpp:qwen4exp/qwen3.8-flash-nextfrom

abdel-darwish-27:qsa-gather-decodeabdel-darwish-27/llama.cpp:qsa-gather-decodeCopy head branch name to clipboard