LLM Digest
Subscribe

Story

arxiv_cs_cl · Nov 17, 2026 · paper

Source brief

Storing the Gemma 4 K=V Cache Once: Exact Global-KV Memory Savings in a Paged Serving Engine with Unchanged Attention Kernels

arxiv.orgNov 17, 2026
original source linked

In brief

In the global attention layers of Gemma 4 31B, the value is the normalized key projection, and the key is the value scaled by a fixed weight, with rotary position encoding on one quarter of its dimensions. The Gemma 4 report states that this design reduces the global KV cache by

Continue reading

Read the original at arxiv.org →Open in live feedRead that day’s brief

Earlier in this thread 4 items