https://github.com/ggml-org/llama.cpp/pull/21038 Since now cache quantization has better quality, does that mean Q8 cache is a good choice now? For example for 26B Gemma4?
