I don't have a chart handy, but in our tests, TRT was ca. 10% faster (but it's much more difficult to setup, you have to convert models, etc). LMDeploy we tested months ago, maybe it's improved now, but they were making fairly wild claims that we didn't observe.
Basically, you shouldn't publish a benchmark without providing all of the details. How long did test run? How were requests staggered? What were the input/output token sizes?
A lot of these 256 in / 256 out benchmarks are useless. If your system prompt is 4k tokens, prefill becomes a serious issue. How performant is your continuous batching? Can you do chunked prefill? Are you running prompt prefix caching (and if so, how performant is that)? It's a lot more complex than simply generating some tokens at various batch sizes.