논문
- WarpQuant: Dual-Domain LLM Quantization via Hadamard Rotation and Output-Fisher Recovery
Harim Choi · ICLR 2027 제출
INT2·INT3 PTQ SOTA(Llama-3 8B INT3 WikiText-2 6.99, VPTQ 7.05), GPTQ W3 대비 디코딩 7.36배인 llama.cpp CUDA 커널
Harim Choi · ICLR 2027 제출
INT2·INT3 PTQ SOTA(Llama-3 8B INT3 WikiText-2 6.99, VPTQ 7.05), GPTQ W3 대비 디코딩 7.36배인 llama.cpp CUDA 커널