fix(container): update image ghcr.io/ggml-org/llama.cpp (server-cuda-b9843 ➔ server-cuda-b9853)

[drag0n141-bot]

This PR contains the following updates:

Package	Update	Change
ghcr.io/ggml-org/llama.cpp	patch	server-cuda-b9843 → server-cuda-b9853

---

Configuration

📅 Schedule: (in timezone Europe/Berlin)

• Branch creation
  • At any time (no schedule defined)
• Automerge
  • At any time (no schedule defined)

🚦 Automerge: Disabled by config. Please merge this manually once you are satisfied.

♻ Rebasing: Whenever PR becomes conflicted, or you tick the rebase/retry checkbox.

🔕 Ignore: Close this PR and you won't be reminded about these updates again.

---

• If you want to rebase/retry this PR, check this box

---

This PR has been generated by Mend Renovate.

[claude]

ghcr.io/ggml-org/llama.cpp: server-cuda-b9843 → server-cuda-b9853

Verdict: Safe to merge

Small patch bump (10 build numbers) in llama.cpp's rolling b#### release scheme. Reviewed every intermediate commit between b9843 and b9853 via the GitHub compare API and per-tag release notes:

• ggml-webgpu: add NVFP4 support (#25143) — WebGPU backend, not used (CUDA image).
• HIP: hipBLAS tuning for gfx900 (#24588) — HIP backend, not used.
• vulkan: matmul roll-bk fix for Asahi Linux (#24663) — Vulkan backend, not used.
• CUDA: fix Gemma E4B MTP FlashAttention (#25148) — model-specific to Gemma E4B MTP; this deployment runs Qwen3.5-4B.
• CUDA: fix get_rows_back grid-y clamp for >65535 rows (#25103) — backward/training-only op, not exercised by the server binary.
• common,server: bracket IPv6 literals in URL authority/logs (#25140) — cosmetic/log-formatting fix; server binds 0.0.0.0, not IPv6.
• model: register t_layer_inp for qwen3next, fixing a crash in speculative-decoding drafters (DFlash/EAGLE3) (#25141) — Qwen3.5's hybrid architecture is qwen3next-derived, but this compose file doesn't configure a draft model / speculative decoding, so the previously-crashing path isn't in use.
• CUDA: fix integer truncation/overflow in flash_attn_mask_to_KV_max KQ mask strides, relevant only at very large context/ubatch sizes (#24945) — this deployment sets LLAMA_ARG_CTX_SIZE=8192, well below the range where the overflow could trigger.
• opencl: initial q1_0 support (#25160) — OpenCL backend, not used.
• ui: remove PWA navigate fallback (#25174) — the deployment sets LLAMA_ARG_NO_UI=1.

No breaking changes or deprecations in this range, and none of the fixes intersect with how docker/nas01/llama/docker-compose.yaml is configured (CUDA backend, no UI, no speculative decoding, 8192 context, non-Gemma model). Only file changed is the image tag bump.

Sources consulted:

• ggml-org/llama.cpp@b9843...b9853
• https://github.com/ggml-org/llama.cpp/releases/tag/b9844 through b9852
• ggml-org/llama.cpp#25141
• ggml-org/llama.cpp#24945
• ggml-org/llama.cpp#25103
• ggml-org/llama.cpp#25140