[P/D] layerwise connector supports DeepSeek-V3.2 sparse attention && Distribute transfer tasks to redundant kv_head cards#5722

wangxiyuan merged 7 commits into

vllm-project:mainvllm-project/vllm-ascend:mainfrom

nwpu-zxr:support_dsv32nwpu-zxr/vllm-ascend:support_dsv32Copy head branch name to clipboard