MiniMax M2.7 GGUF Investigation, Fixes, Benchmarks
An investigation into MiniMax-M2.7 GGUF revealed that perplexity NaNs affect 21-38% of GGUFs on Hugging Face. The issue was traced to overflowing in llama.cpp, specifically in `blk.61.ffn_down_exps` for Q5_K and Q4_K quantizations, and the team has fixed theirs.
