KVQuant HTML: nuq3-1% perplexity, passkey, and A6000 kernel latency (arXiv:2401.18079v6)

HTML of arXiv:2401.18079v6, read 2026-10-18. Distinct from pap-quant-kvquant (the /abs page) because the numbers live in the HTML tables. Table 1 prose: nuq4-1% <0.02 PPL, nuq3-1% <0.1 PPL, nuq2-1% <0.5 PPL vs fp16 on WikiText-2 across LLaMA models, with 3.7x / 4.8x / 6.9x memory savings. Table 2: LLaMA-2-7B-32K passkey success nuq3-1% is 0.98/1/1/1/1 at 2K-32K vs fp16 1.0, avg bit width 3.33. Table 3: LongBench average 31.21 (KVQuant-3bit-1%) vs 31.96 fp16 vs 30.04 KIVI-2. Table 22: A6000 kernel latency for LLaMA-2-7B-32K nuq4-1% vs fp16 matvec, key total 126.3 us vs 219.4 us at 16K. Abstract: LLaMA-7B up to 1M context on one A100-80GB and 10M on 8 GPUs. Appendix Table 8: LLaMA-7B fp16 KV is 64.0 GB at 128K and 512.0 GB at 1M; nuq3 is 12.0 / 96.1 GB.