ncnn has a pipeline cache for vulkan pipeline creation. It avoids creating the same shader module, descriptor set layout, pipeline layout, compute pipeline and descriptor update template repeatedly ...
echo "[1/10] Running llama2 with harmbench-test50, kv_cache=None" bash ./eval/llama2.sh --src-path ../src --eval-cfg 1 --dataset harmbench-test50 --attack gcq --kv ...
一些您可能无法访问的结果已被隐去。
显示无法访问的结果