prefix caching on repeated workflows

[tensorriver]

Posting this in case it saves someone else a little time.

I'm not trying to build a perfect architecture diagram. I just want a sane starting point around vLLM and 7900 XTX.

Would rather learn from your scars than collect my own all at once.

[vramscope]

That matches what I was seeing. The pattern around quant drift has been pretty similar for me.

[cacheline]

A couple questions before I guess too hard. Are you testing this on mini PC or something else?