vllm

Deepseek optimizations [experimental duplicate of https://github.com/vllm-project/vllm/pull/28540]

#29497

Open