inference performance

TurboQuant vs FP8: vLLM’s first broad test

vLLM found FP8 KV-cache quantization beats TurboQuant on speed, while TurboQuant’s strongest variants hurt accuracy.