Feature Request: Please merge upstream PR #22288 for Gemma4 SWA cache reuse fix

[Heritager-tech]

Prerequisites

• I am running the latest code. Mention the version if possible as well.
• I carefully followed the README.md.
• I searched using keywords relevant to my issue to make sure that I am creating a new issue that is not already open (or closed).
• I reviewed the Discussions, and have a new and useful enhancement to share.

Feature Description

Gemma4 models don't reuse prompt cache due to missing PR ggml-org#22288 from upstream llama.cpp:
ggml-org#22288

Current behavior: "forcing full prompt re-processing due to lack of cache data"
Expected: checkpoint restoration works with --swa-full

please sync with upstream master to include this fix

Motivation

Official llama.cpp b9127 works perfectly - cache reuse gives speedup for repeated prompts

Possible Implementation

No response

[nnnarvaez]

Seeing repetitions on Gemma4 26B MoE (specially in reasoning CoT) As mentioned here: ggml-org#22288
I have been searching all around both repos for something similar.

Had Gemini check the llama-server logs and it pointed to the SWA cache not being reused (thou i don't see a relation) It advised to run with --swa-full + -fa as mitigation measure but the vRAM requirements skyrocketed.

My Toaster:

ggml_cuda_init: found 2 CUDA devices (Total VRAM: 24569 MiB):
  Device 0: NVIDIA GeForce RTX 4070 SUPER, compute capability 8.9, VMM: yes, VRAM: 12281 MiB
  Device 1: NVIDIA GeForce RTX 3060, compute capability 8.6, VMM: yes, VRAM: 12287 MiB
version: 9082 (5aeb2fdbe)
built with GNU 13.3.0 for Linux x86_64