A personal blog and about me.
Oct 16, 2026
I deployed Qwen3.8-27B (NVFP4) on a single NVIDIA DGX Spark and benchmarked vLLM+MTP against SGLang+DSpark. The honest result: one number is a lie, and there's a memory leak you need to know about.