latency vs throughput for small deployments on 7900 XTX

[cudaworks]

My first guess would be this: Before rebuilding anything, I'd test one clean baseline with Mistral and compare from there. Once you know the plain pat

that fixed it for me too.