page attention does not work

[chenglu66]

i dont konw when this happen when batch increasae generation throughput: did not increase as i think
when I use VLLM
batch 1 Avg generation throughput: 55.2 tokens/s
batch 3 Avg generation throughput: 98.2 tokens/s,
batch 5 Avg generation throughput: 134.4 tokens/s
batch 30 Avg generation throughput: 249.2 tokens/s
batch 256 Avg generation throughput: 1275.0 tokens/s