CUDA update broke my throughput

[edgecoast]

Interesting, I hadn't tried that. I ended up in about the same place using SGLang.

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

If you post your exact settings, people here can probably narrow it down pretty fast.