Gemma 4 QAT 31B responds better to KV cache quantization too

Gemma 4's quantization-aware training (QAT) model shows improved performance with key-value (KV) cache quantization, particularly with a Q8_0 configuration. Benchmark results using KL Divergence on Wikitext with a 16k context indicate that the model maintains a 99.9% KLD, suggesting effective attention retention on important tokens. This improvement is significant for practitioners as it enhances the efficiency of deploying LLMs in resource-constrained environments while maintaining performance.