Better perplexity for 2- and 3-bit quantization for LLaMA-v2-70B#2807

ikawrakow merged 2 commits intomasterggml-org/llama.cpp:masterfrom

ik/refine_70Bggml-org/llama.cpp:ik/refine_70BCopy head branch name to clipboard