prefill bottlenecks on longer prompts with vLLM

[northquill]

My first guess would be this: Before rebuilding anything, I'd test one clean baseline with Mistral and compare from there. Not glamorous advice, but repeatable baselines save a lot of time.