Papers
- WarpQuant: Dual-Domain LLM Quantization via Hadamard Rotation and Output-Fisher Recovery
Harim Choi · Submitted to ICLR 2027
INT2·INT3 PTQ SOTA (Llama-3 8B INT3 WikiText-2 6.99 vs VPTQ 7.05) and a llama.cpp CUDA kernel that decodes 7.36× faster than GPTQ W3