prefix caching on repeated workflows with TensorRT-LLM

[cacheline]

I figured I'd ask here before changing everything.

Good point. I don't think your conclusion is far off. The biggest change for me was simplifying the stack instead of adding one more optimization layer.