SGLang, Qwen, and NVIDIA collaborate on NVFP4 KV cache for long context

SGLang integrated NVIDIA's NVFP4 4-bit float format into its KV cache, reducing memory usage to 56.25% of FP8 while preserving accuracy. Tests showed up to 78% higher decode throughput for long contexts and significant benefits for agentic workloads due to higher cache hit rates.