# WUSH-KV

The work adapts the WUSH transform, originally for weight-activation quantization, to the KV cache of grouped-query attention: calibration data builds separate key and value transforms per KV head (the value transform folded into the weights, the key transform applied after RoPE), the transform is shown to be near-optimal among sensitivity-balanced transforms for the QuEST quantizer, and end-to-end evaluation in SGLang with OSCAR-style percentile-clipped affine quantization gives 2-bit WikiText-2 perplexity of 10.51 on Qwen3-8B versus OSCAR's 13.74, with the 8B model scoring higher than OSCAR on AIME 2025, MATH-500, GPQA Diamond, and LiveCodeBench v6.

The work adapts the WUSH transform, originally for weight-activation quantization, to the KV cache of grouped-query attention: calibration data builds separate key and value transforms per KV head (the value transform folded into the weights, the key transform applied after RoPE), the transform is shown to be near-optimal among sensitivity-balanced transforms for the QuEST quantizer, and end-to-end evaluation in SGLang with OSCAR-style percentile-clipped affine quantization gives 2-bit WikiText-2 perplexity of 10.51 on Qwen3-8B versus OSCAR's 13.74, with the 8B model scoring higher than OSCAR on AIME 2025, MATH-500, GPQA Diamond, and LiveCodeBench v6.