https://hgpu.org/?p=27120
LLM.int8(): 8-bit Matrix Multiplication for Transformers at Scale