throughput per watt reality check for Qwen

[kernelmike]

I hit almost the same issue last night.

Appreciate the detail. 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.