KVarN: Native vLLM KV-cache quantization back end by Huawei

Huawei released KVarN, a native KV-cache quantization back end for vLLM that delivers up to 5x more cache capacity and 1.3x the throughput of FP16 while maintaining FP16-level accuracy. The calibration-free system handles agentic and long-context workloads by quantizing keys at 4 bits and values at 2 bits, outperforming existing methods like TurboQuant by up to 2.4x in throughput with higher accuracy. KVarN ships as a vLLM fork that requires only a single flag change to enable, eliminating the need for model modifications or calibration data.

⚡️ Built for agentic and long-context workloads. 💡 KVarN delivers 3-5x more KV-cache capacityandup to ~1.3x the throughputof FP16, so you fit far longer contexts and serve more concurrent requests, withFP16-level accuracy. 🔌 Calibration-free, plug-and-play with vLLM.A native vLLM attention backend: add one flag, no model changes, no calibration. 🥊 Up to ~2.4× TurboQuant throughput, same capacity,higher accuracy. kvarn/kvɑːɳ/ ·noun Swedish - A grinding apparatus used to reduce substances into smaller particles or powder, especially grains, seeds, spices, coffee beans, KV-caches. KV-cache quantization usually comes with a catch. As the vLLM TurboQuant blog https://vllm.ai/blog/2026-05-11-turboquant shows, existing methods buy extra KV-cache capacity but give up throughput TurboQuant reports 40 to 52% lower throughput for 2.3-3.7x capacity , and aggressive low-bit quantization also tends to cost accuracy . Losing both speed and quality is the main reason KV-cache quantization is rarely turned on in production. KVarN is built to keep both. On Qwen3-32B AIME25, 16K-context burst, TP=2 it matches FP16 accuracy and beats its throughput while delivering ~4× the KV-cache capacity: KVarN stays in the upper-right corner the blog's methods can't reach: FP16-level accuracy, FP16-or-better throughput, and several times the context. KVarN ships as a vLLM fork. Install it like vLLM, then select the KVarN KV-cache dtype. 1. Clone git clone https://github.com/huawei-csl/KVarN.git cd KVarN 2. Install uses the upstream precompiled wheel; KVarN kernels are Triton, JIT-compiled at runtime VLLM USE PRECOMPILED=1 pip install -e . python from vllm import LLM, SamplingParams llm = LLM model="Qwen/Qwen3-32B", dtype="float16", KVarN runs in float16 kv cache dtype="kvarn k4v2 g128", enable KVarN block size=128, KVarN tile size print llm.generate "Explain KV-cache quantization in one sentence.", SamplingParams max tokens=64 0 .outputs 0 .text Serving works the same way: vllm serve Qwen/Qwen3-32B --dtype float16 --kv-cache-dtype kvarn k4v2 g128 --block-size 128 Note:KVarN runs in float16 compute. The tile / page size is currently fixed at 128 one vLLM block = one KVarN tile ; other page sizes are coming soon. Tip capacity :KVarN realizes its full KV-cache capacity when there is room to amortize a small fixed decode workspace. On multi-GPU or generous --gpu-memory-utilization setups this is automatic. On a tight single-GPU budget, vLLM's CUDA-graph memory profiler can over-reserve and shrink the KV pool; set VLLM MEMORY PROFILER ESTIMATE CUDAGRAPHS=0 and/or raise --gpu-memory-utilization to recover the full capacity. KVarN quantizes the KV cache one fixed-size token tile at a time, walking each tile through the four stages above: - Cache : the raw fp16 KV tile channels × tokens , straight from attention. - Rotated Cache : a Hadamard rotation along the channel dimension mixes channels so that per-channel outliers are spread out, making the tile easier to quantize. The rotation is orthonormal, so attention scores are preserved. - Normalized Cache : iterative variance normalization Sinkhorn-like alternates column- and row-wise standard-deviation normalization in log space, equalizing variance across the tile and shrinking quantization error before any rounding happens. - Quantized Cache : asymmetric round-to-nearest at low bit-width, with the scales folded back in at read time keys per channel, values per token . The shipped preset spends more bits on keys than values kvarn k4v2 g128 : 4-bit keys, 2-bit values . We chose to release this configuration because it meets the strictest accuracy bar, matching FP16, that the most demanding production deployments and vLLM require, while still delivering throughput above FP16. KVarN is the official vLLM implementation of our paper: 📄 KVarN: Variance-Normalized KV-Cache Quantization Mitigates Error Accumulation in Reasoning Tasks arXiv:2606.03458 If you use KVarN, please cite: @misc{muller2026kvarn, title={KVarN: Variance-Normalized KV-Cache Quantization Mitigates Error Accumulation in Reasoning Tasks}, author={Lorenz K. Muller and Philippe Bich and Chiara Boretti and Hyun-Min Chang and Jiawei Zhuang and Lukas Cavigelli}, year={2026}, eprint={2606.03458}, archivePrefix={arXiv}, primaryClass={cs.LG}, url={http://arxiv.org/abs/2606.03458} } KVarN is built on vLLM https://github.com/vllm-project/vllm v0.22.0 and is released under the Apache 2.0 License. The original vLLM README is preserved as README vLLM.md /huawei-csl/KVarN/blob/main/README vLLM.md .