vLLM Enables FlashInfer for Gemma 4 with NVFP4 KV Cache
November 13, 2026
vLLM now supports using FlashInfer for Gemma 4 when an NVFP4 KV cache is used and FA4 is unavailable. This provides a high-performance fallback for inference.
HOW THIS AFFECTS YOU
●
builderYou can maintain high throughput for Gemma 4 using NVFP4 even without FA4 support.