SGLang, Qwen, NVIDIA implement NVFP4 KV cache: decode throughput up 78% at 4-bit

SGLang collaborated with Qwen and NVIDIA to support NVFP4 format for KV cache quantization in the SGLang inference engine. This reduces KV cache memory to ~56% of FP8, enabling larger batch sizes and higher hit rates in radix cache. For 1M context, decode throughput improves up to 78% in capacity-bound scenarios. Accuracy loss is minimal except for -1.6 pp on SWE-bench.