Was surprised to see no results for Laguna S 2.1 on DGX spark on localmaxxing.com/en, so I submitted what I get with the official recipe on my end. (Hello World @LottoLabs) vLLM NVFP4+DFlash: 21.6 tok/s decode, ~2300 prefill tok/s llama.cpp Q4_K_M: 19.3 tok/s decode, ~790 prefill tok/s NVFP4 safetensors on vLLM: Q4_K_M GGUF: localmaxxing.com/en/models/poolside/Laguna-S-2.…