# llama.cpp (llama-server) In-cluster LLM inference via llama.cpp's `llama-server`, serving a local model on the NUCBox APU (AMD Ryzen AI Max 395 / Strix Halo, Radeon 8060S, ~120 GiB unified memory: ~90 GiB VRAM / 30 GiB CPU RAM). LiteLLM (`litellm/`) points at these in-cluster Services instead of the external `10.88.20.12:11434` Ollama endpoint. ## Layout One Deployment + Service **per model**, all in namespace `llamacpp`, all pinned to the NUCBox (`nodeSelector: {kubernetes.io/arch: amd64, hardware: high-memory}`): | Alias | Model | GGUF | Service | Args ref | |--------------------------|--------------------------------|-------------------------------------------------------|------------------------------------------------------|----------| | `deepseek-v4-flash-0731` | DeepSeek-V4-Flash-0731 (MoE) | unsloth/DeepSeek-V4-Flash-0731-GGUF (UD-IQ1_M, ~87 GiB) | `llamacpp-deepseek-v4-flash-0731.llamacpp:80` | [args-deepseek-v4-flash-0731.md](args-deepseek-v4-flash-0731.md) | DeepSeek-V4-Flash-0731 is a Mixture-of-Experts model (256 experts, 6 active per token) with MLA attention, so only a small fraction of the weights is computed per token. The full ~87 GiB of IQ1_M weights is loaded into the unified memory pool. > **Previously** the NUCBox ran two co-resident Qwen3.6 models (a 27B dense and > a 35B-A3B MoE "flash"). DeepSeek-V4-Flash-0731 (~87 GiB) nearly fills the > whole 90 GiB VRAM pool on its own, so both Qwen models were removed to make > room. Their GGUF files are deleted from the shared PVC by the new pod's > `fetch-model` initContainer on first boot. Their deployment arguments are > still documented for redeployment: > - [args-qwen36-27b.md](args-qwen36-27b.md) — dense 27B (deeper reasoning) > - [args-qwen36-35b-a3b.md](args-qwen36-35b-a3b.md) — MoE 35B-A3B "flash" (fast) Model files are downloaded idempotently by an initContainer into a shared hostPath PVC (`/data/llamacpp/models` on the NUCBox), so pods survive reboots without re-downloading. The UD-IQ1_M GGUF is split across 3 shards (`-00001-of-00003` … `-00003-of-00003`); llama.cpp auto-loads all shards when pointed at the first one. ## GPU / Vulkan The `server-vulkan` image (`ghcr.io/ggml-org/llama.cpp:server-vulkan`) bundles the Mesa/RADV Vulkan driver, which supports the Radeon 8060S (RDNA 3.5). The Vulkan backend supports the `IQ1_M` matmul (including the MoE `matmul_id` variant), so the whole model runs on the GPU. `deepseek4` is a brand-new arch (2026-07), so the floating `server-vulkan` tag is used to pull a recent enough build — pin to a specific `server-vulkan-bXXXX` tag once a known-good one is verified. The container mounts `/dev/dri` and runs `privileged: true` — the simplest reliable way to give Vulkan access to the DRM render node on k3s without a device plugin. Tighten later with `supplementalGroups` (the host's `render` group GID) if desired. ### Verify the GPU is actually used ```bash kubectl logs -n llamacpp deploy/llamacpp-deepseek-v4-flash-0731 | grep -iE 'vulkan|gpu|offload|device' ``` If only a CPU device shows up, the container can't see the GPU — check that `/dev/dri/renderD128` exists on the NUCBox and that the `amdgpu` module is loaded. ## VRAM budget (single model) The model (~87 GiB IQ1_M) is almost the size of the entire 90 GiB VRAM pool, so it **cannot be fully offloaded**: `-ngl 999` would try to put all 43 layers into VRAM and overflow once the KV cache + Vulkan compute buffers are added. Instead `-ngl 40` offloads 40 of 43 layers to the GPU and keeps the last 3 (~6 GiB) on CPU RAM, leaving ~8 GiB of VRAM headroom for the KV cache, compute buffers, and co-resident pods. Approximate VRAM usage: | Component | VRAM | |---------------------------------|-------------| | Weights (40 GPU layers) | ~81 GiB | | KV cache (q8_0, 64k, 1 slot) | ~1.6 GiB | | Vulkan compute buffers | ~2 GiB | | **Total in VRAM** | **~85 GiB** | | **Headroom (of 90 GiB)** | **~5–8 GiB**| 3 layers (~6 GiB) live in CPU RAM (counted against the pod's cgroup memory limit, not VRAM). KV cache is tiny thanks to DeepSeek-V4's **MLA** attention (`num_kv_heads=1`, `head_dim=512` + 64 decoupled RoPE ⇒ ~576 elements/token/layer). At 64k context, q8_0 KV is only ~1.6 GiB, so context is cheap — `-c` is capped at 65536 (the required minimum) to maximise VRAM headroom, not because KV is the constraint. ## Tuning The key knobs (in `deployment-deepseek-v4-flash-0731.yaml`): - `-ngl 40` — offload 40 of 43 layers to GPU. The model (~87 GiB) is nearly the whole 90 GiB VRAM pool, so full offload would overflow once KV cache + compute buffers are added. Keeping 3 layers (~6 GiB) on CPU leaves ~8 GiB headroom. Raise toward 43 if VRAM allows; lower (e.g. 38) if the pod OOMs / Vulkan runs out of device memory. - `-c 65536` — total KV-cache context (64k, the required minimum). 1 slot gets the full 64k. MLA KV is tiny (~1.6 GiB at q8_0), so context is cheap; capped at the minimum to maximise VRAM headroom. Raise if headroom allows. - `-np 1` — 1 parallel slot (the full 64k goes to a single concurrent request). Extra slots multiply KV VRAM (cheap here), but 1 slot keeps headroom maximal. - `--cache-type-k q8_0 --cache-type-v q8_0` — quantize the KV cache to q8_0, halving KV VRAM with ~negligible quality loss. Essential to keep headroom. - `--temp 1.0 --top-p 0.95` — default sampling parameters (DeepSeek-V4 recommendation). These are server defaults; clients can override per request via the OpenAI-compatible API. - `--threads 8` — CPU threads for sampling + the 3 CPU-resident layers. ## Memory accounting k8s sees only the ~30 GiB system RAM as allocatable (the ~90 GiB VRAM is reserved by firmware and managed by `amdgpu`). The GPU-resident model weights and KV cache live in VRAM and are **not** counted against the container's cgroup memory limit — that limit only covers CPU-side overhead, the mmap'd GGUF pages for the 3 CPU-resident layers (~6 GiB), and reclaimable page cache during load. If the pod is OOM-killed during model load, raise the memory limit (and/or lower `-ngl`). ## Adding / replacing a model 1. Copy `deployment-deepseek-v4-flash-0731.yaml` → `deployment-.yaml`; change the `model:` label, GGUF URL/file(s), `--alias`, and Service name. For split GGUFs, point `-m` at the first shard and download all shards in the `fetch-model` initContainer. 2. Point LiteLLM at it in `litellm/litellm.yaml`: ```yaml - model_name: litellm_params: model: openai/ api_base: http://.llamacpp/v1 api_key: "sk-no-auth" ``` 3. (No gen-apps.sh change needed — the `llamacpp` app already syncs the whole directory recursively.) 4. Check the VRAM budget table above — at ~87 GiB this model nearly fills the 90 GiB pool on its own, so co-locating another large model is not possible.