Qwen3-8B on workstation Blackwell: vLLM vs SGLang vs llama.cpp, plus an FP8 pass

Benchmarks of the same model on the same GPU across three serving stacks, then an FP8 pass on the...