The document states that up to 3.5x higher throughput than TGI，I have my doubts

[Storm0921]

What is the TGI version you are comparing? This is not consistent with the benchmark results of some friends，I'm a bit confused.

https://zhuanlan.zhihu.com/p/645732302

[LiuXiaoxuanPKU]

TGI numbers are before adding paged attention to their repo. The performance might be similar after TGI adds vllm's techniques. (after this pr)