Optimize DeepSeek V4 qkv_proj_rope decode (S=2): partial-sum reduces, amax fold, K-tile/stage tuning#339

zhangqi-chen merged 2 commits into

hw-native-sys:mainhw-native-sys/pypto-lib:mainfrom

wangqin1723-max:perf/dsv4-qkv-proj-rope-grouped-chunkingwangqin1723-max/pypto-lib:perf/dsv4-qkv-proj-rope-grouped-chunkingCopy head branch name to clipboard