LLM inference optimization, DGX Spark, SGLang, Qwen, NVFP4, FP8, speculative decoding, long-context inference