initialize --model pagedattention

PagedAttention

Memory-efficient inference at scale

# KV cache management for large language models
Non-contiguous memory allocation. Zero fragmentation.
Page-level granularity. O(1) complexity.

● vLLM compatible   ● throughput +2-4x   ● memory -60%