throughput per watt reality check for Qwen

[deltapoint17]

Spent some time on this today and wanted to compare notes.

I've gone back and forth between vLLM and llama.cpp for Llama. The interesting difference for me has been token cost.

Trying to avoid rebuilding the whole thing twice.