vllm

[Core] Use individual MM items in P0/P1 cache and model runner

#22570

Merged