Name and Version

llama-server: b8648

Operating systems

Linux

GGML backends

CUDA

Hardware

RTX 5090 32GB vram

Models

Gemma4-31B-it-Q5-K-M

Problem description & steps to reproduce

Context shift isn't working when using Gemma-4. kv-quantization with context-shift isn't working aswell. :
using latest build of llamacpp rn.

First Bad Commit

No response

Relevant log output

slot update_slots: id 3 | task 102 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)