Should I be seeing more of a performance leap when using NVFP4, INT4, FP8 with VLLM over MXFP4, Q4, and Q8 with llama.cpp based inference on Blackwell based GPUs?
I hope I am doing something wrong here but I am seeing about almost double the t/s using LM studio with Qwen3.5 and Nemotron models than I am seeing with Nvidia’s own vLLM containers built for Spark. I was surprised I was only getting 15-ish t/s with N…