Per-IC quantization offers an effective solution to isolate activation outliers in large language models, leading to significant improvements in performance.
