88plug · Oct 29, 2026 · Oct 29, 2026 · Oct 29, 2026 · Oct 29, 2026
diff --git a/README.md b/README.md
@@ -944,7 +944,7 @@ OpenAI API at `/v1` from `svc/vllm`:
 
 | Backend | Image (pinned by digest) | Notes |
 |---------|--------------------------|-------|
-| `vllm` | `vllm/vllm-openai-xpu:v0.31.0` | upstream vLLM for Intel GPUs; validated on Arc Pro B-series |
+| `vllm` | `vllm/vllm-openai-xpu:v0.32.0` | upstream vLLM for Intel GPUs (Data Center GPU and Arc, per [vLLM's XPU guide](https://github.com/vllm-project/vllm/blob/v0.32.0/docs/getting_started/installation/gpu.xpu.inc.md)); v0.32.0 not yet run here on an Arc B-series card |
 | `llm-scaler` | `intel/llm-scaler-vllm:1.5` (vLLM 0.26.0-b2) | Intel's build for Arc B60/B70 |
 
 ```bash
@@ -1193,7 +1193,7 @@ and submits through execlists, so a GuC fault there warns instead. A failed HuC
 only warns: it runs media decode and encode, not inference. Without root the GuC is reported as not checked. The refused GPU's
 facts/v1 target reads `usable: false` with the `reason` and `fix`; `plan` then
 falls back to the next green cell on a usable target (OVMS on the CPU, or
-llama.cpp on Vulkan, pinned as `ghcr.io/ggml-org/llama.cpp:server-vulkan-v0.5.0`
+llama.cpp on Vulkan, pinned as `ghcr.io/ggml-org/llama.cpp:server-vulkan-b11960` (release v0.8.0)
 by digest) and lists the fix. With nothing left to pick the decision is `fix`,
 not an engine; a Gaudi without its driver plans the same way. `agent up` follows
 the same verdict: `auto` uses OVMS on the CPU and says why, and

diff --git a/bench/competitors.lock b/bench/competitors.lock
@@ -1,6 +1,6 @@
 {
   "schema": "k3d-gpu.competitors-lock/v1",
-  "note": "Competitor stacks for 'k3d-gpu bench --vs'. Every image is pinned by registry digest (verified with docker buildx imagetools inspect on 2026-10-17). Each stack serves the same model family and 4-bit tier as the k3d-gpu arm (Qwen3-4B, int4 / Q4_K_M), with its own vendor defaults. Weights are pinned too: after the model loads, 'bench' hashes artifact_path inside the container and refuses to measure unless it equals artifact_sha256 (the Hugging Face LFS sha256 at model_revision, or, for Ollama, the model-layer blob digest listed in the pinned registry manifest; Ollama rewrites its on-disk manifest copy, so that file does not hash to the registry digest).",
+  "note": "Competitor stacks for 'k3d-gpu bench --vs'. Every image is pinned by registry digest (verified with docker buildx imagetools inspect on 2026-10-17; the llama.cpp and Ollama pins moved to llama.cpp v0.8.0 and Ollama v0.40.13 on 2026-10-29, verified the same way). Each stack serves the same model family and 4-bit tier as the k3d-gpu arm (Qwen3-4B, int4 / Q4_K_M), with its own vendor defaults. Weights are pinned too: after the model loads, 'bench' hashes artifact_path inside the container and refuses to measure unless it equals artifact_sha256 (the Hugging Face LFS sha256 at model_revision, or, for Ollama, the model-layer blob digest listed in the pinned registry manifest; Ollama rewrites its on-disk manifest copy, so that file does not hash to the registry digest).",
   "stacks": {
     "llamacpp": {
       "api": "/v1",
@@ -9,9 +9,9 @@
       "artifact_sha256": "7485fe6f11af29433bc51cab58009521f205840f5b4ae3a32fa7f92e8534fdf5",
       "cache": "/root/.cache/llama.cpp",
       "device": "cpu",
-      "docs": "https://github.com/ggml-org/llama.cpp/blob/b11706/common/arg.cpp",
+      "docs": "https://github.com/ggml-org/llama.cpp/blob/v0.8.0/common/arg.cpp",
       "health": "/health",
-      "image": "ghcr.io/ggml-org/llama.cpp:server-b11706@sha256:f41d47eee678c02656eefc9ed8cce76cab584aa7032805eeef0e6639fcfda485",
+      "image": "ghcr.io/ggml-org/llama.cpp:server-b11960@sha256:8052c13edd987de8858f8150031e025ac3d66823f09cf3ec8220899c4abce220",
       "model": "Qwen/Qwen3-4B-GGUF",
       "model_revision": "Qwen/Qwen3-4B-GGUF@bc640142c66e1fdd12af0bd68f40445458f3869b",
       "port": 8080,
@@ -26,10 +26,10 @@
       "artifact_sha256": "163553aea1b1de62de7c5eb2ef5afb756b4b3133308d9ae7e42e951d8d696ef5",
       "cache": "/root/.ollama",
       "device": "cpu",
-      "docs": "https://github.com/ollama/ollama/blob/v0.40.5/openai/openai.go",
+      "docs": "https://github.com/ollama/ollama/blob/v0.40.13/openai/openai.go",
       "health": "/api/version",
       "ignore_eos": false,
-      "image": "ollama/ollama:0.40.5@sha256:5d3a2f25f9825628daba94dc0b1d421c7d57d6c03ae5faa9ba050e058bab8241",
+      "image": "ollama/ollama:0.40.13@sha256:b2f0122c476bc9fe7ce679d8bb4bd555351e25e2554eefb156f61cb0a89cf429",
       "model": "qwen3:4b-q4_K_M",
       "model_revision": "registry.ollama.ai/library/qwen3:4b-q4_K_M manifest sha256:2bfd38a7daaf4b1037efe517ccb73d1a3bbd4822cf89f1a82be1569050a114e0, model layer sha256:163553aea1b1de62de7c5eb2ef5afb756b4b3133308d9ae7e42e951d8d696ef5",
       "port": 11434,

diff --git a/console/model_opt.py b/console/model_opt.py
@@ -94,9 +94,9 @@
 OVMS_SPEC_DEMO = OVMS_DOC + "demos/continuous_batching/speculative_decoding/README.md"
 OVMS_TOOL_PARSERS = ("llama3", "hermes3", "phi4", "mistral", "gptoss", "qwen3coder", "devstral", "lfm2", "gemma4", "onyx", "minicpm5")
 OVMS_REASONING_PARSERS = ("qwen3", "gemma4", "gptoss", "minicpm5", "lfm2", "onyx")
-VLLM_SPEC_DOC = "https://github.com/vllm-project/vllm/blob/v0.31.0/vllm/config/speculative.py"
-VLLM_MTP_DOC = "https://github.com/vllm-project/vllm/blob/v0.31.0/docs/features/speculative_decoding/mtp.md"
-VLLM_HW_TABLE = "https://github.com/vllm-project/vllm/blob/v0.31.0/docs/features/README.md (hardware table, L67)"
+VLLM_SPEC_DOC = "https://github.com/vllm-project/vllm/blob/v0.32.0/vllm/config/speculative.py"
+VLLM_MTP_DOC = "https://github.com/vllm-project/vllm/blob/v0.32.0/docs/features/speculative_decoding/mtp.md"
+VLLM_HW_TABLE = "https://github.com/vllm-project/vllm/blob/v0.32.0/docs/features/README.md (hardware table, L67)"
 VLLM_GAUDI_SD = "https://github.com/vllm-project/vllm-gaudi/blob/v0.30.0/docs/features/supported_features.md (L25, L27)"
 LLAMACPP_ARGS = "https://github.com/ggml-org/llama.cpp/blob/v0.8.0/common/arg.cpp (L4322 -md)"
 LLAMACPP_SPEC = "https://github.com/ggml-org/llama.cpp/blob/v0.8.0/common/arg.cpp (L4322 --spec-draft-model, L4330 --spec-type)"
@@ -709,7 +709,7 @@ def optimize(bundle, engine, device, opts=None):
     o = dict(opts or {})
     thinking_want = o.get("thinking") or "off"
     draft_want = o.get("draft") or "auto"
-    vllm_version = tuple(int(x) for x in (o.get("vllm_version") or "0.31.0").split(".")[:3])
+    vllm_version = tuple(int(x) for x in (o.get("vllm_version") or "0.32.0").split(".")[:3])
     prefix_caching = o.get("prefix_caching")
     hw_class = o.get("hw_class")
     target = bundle["target"]
@@ -976,7 +976,7 @@ def choose(fact, choice, source):
             doc["text_only"] = True
             doc["engine_args"].append("--language-model-only")
             choose("vision%s tower in config.json; agents send text" % ("/audio" if facts["audio"] else ""),
-                   "--language-model-only (no encoder memory, larger KV pool)", "https://github.com/vllm-project/vllm/blob/v0.31.0/vllm/engine/arg_utils.py (L1429, v0.17.0+)")
+                   "--language-model-only (no encoder memory, larger KV pool)", "https://github.com/vllm-project/vllm/blob/v0.32.0/vllm/engine/arg_utils.py (L1460, v0.17.0+)")
         elif engine == "ovms":
             choose("vision tower in the IR", "loaded as a VLM pipeline: OVMS refuses LM on a VLM directory",
                    OVMS_DOC + "src/llm/servable_initializer.cpp (L588-641)")

diff --git a/console/moe_predict.py b/console/moe_predict.py
@@ -55,6 +55,8 @@
 
 # Engines per weight container, as the H54 engine scout verified them at the
 # pinned versions (EXPERIMENTS.md H54 pre-registration). Order is preference.
+# llama.cpp names the agent's own llama.cpp pins (release v0.8.0, tag b11960),
+# newer than the b11749 build the scout read.
 ENGINES = [
     {"name": "ovms", "container": "ov", "devices": ("cpu", "gpu"), "tp": False, "formats": None,
      "images": ("openvino/model_server:2026.4.1@sha256:148e3fe802261ed0b5b634de3bbccdefe7176a218ec34a88e382b40704c34a1c",
@@ -66,8 +68,8 @@
     {"name": "sglang-cpu", "container": "safetensors", "devices": ("cpu",), "tp": False,
      "formats": ("bf16", "fp8", "w4a16"), "images": ("lmsysorg/sglang:v0.5.22-xeon@sha256:4e5c413ed9a31f38f74e18aa9b645d65a0b0a023d42f63edb4e2b412ea1a230f",)},
     {"name": "llama.cpp", "container": "gguf", "devices": ("cpu", "gpu", "hybrid"), "tp": False,
-     "formats": None, "images": ("ghcr.io/ggml-org/llama.cpp:server-b11749@sha256:9fbaa06d7f4449e2ae0d02b40e0431f7c212dcf231c9e3a7777a6afc2cc8f312",
-                "ghcr.io/ggml-org/llama.cpp:server-intel-b11749@sha256:ae2d41ba73574a1e12cd237c88b512c8bb39ba6f703e873912b76cb8988a81bd")},
+     "formats": None, "images": ("ghcr.io/ggml-org/llama.cpp:server-b11960@sha256:8052c13edd987de8858f8150031e025ac3d66823f09cf3ec8220899c4abce220",
+                "ghcr.io/ggml-org/llama.cpp:server-intel-b11960@sha256:068cd41cdc292bed89b4a6b7740ea7a33df8ba26bf880d0e12ac75972dce1700")},
 ]
 # Architectures with a narrower engine set than their container suggests.
 ARCH_ENGINES = {"Qwen4ExpForConditionalGeneration": ("llama.cpp",), "qwen4exp": ("llama.cpp",)}

diff --git a/console/plan.py b/console/plan.py
@@ -255,7 +255,7 @@ def plan(facts, matrix, overrides=None):
     return doc
 
 
-AGENT_VLLM_VERSION = {"HPU": "0.30.0", "GPU": "0.31.0"}  # the agent's pinned vLLM images
+AGENT_VLLM_VERSION = {"HPU": "0.30.0", "GPU": "0.32.0"}  # the agent's pinned vLLM images
 
 
 def agent_target(facts, doc, backend=None):
@@ -312,7 +312,7 @@ def add_optimizer(doc, facts, meta, opts):
     o = dict(opts or {})
     backend = o.pop("backend", None)
     engine, device, t = agent_target(facts, doc, backend)
-    o.setdefault("vllm_version", "0.26.0" if backend == "llm-scaler" else AGENT_VLLM_VERSION.get(device, "0.31.0"))
+    o.setdefault("vllm_version", "0.26.0" if backend == "llm-scaler" else AGENT_VLLM_VERSION.get(device, "0.32.0"))
     o["hw_class"] = hw_class(t)
     # No backend set: as `agent up`, the CPU engine follows the ledger (model_opt.py H59).
     o.setdefault("engine_from_ledger", backend is None)

diff --git a/docs/intel/upstream-eat.md b/docs/intel/upstream-eat.md
@@ -51,7 +51,7 @@ targets #19. What it changes for a k3d or k3d-gpu cluster, read from source only
 
 | #21 adds | Already covered by k3d-gpu | Expected on a k3d-gpu cluster |
 |---|---|---|
-| Arc B60/B70 XPU path (`compute_platform=xpu`, `vllm-openai-xpu:v0.28.0`, TP4 `gemma-4-31B-it`) | B70 detection, `gpu.intel.com/xe` asserted by `up`, vLLM XPU `v0.31.0` by digest through `agent up` | Host Level Zero and OpenCL libraries mounted by `hostPath` resolve inside the k3d node container, not on the host |
+| Arc B60/B70 XPU path (`compute_platform=xpu`, `vllm-openai-xpu:v0.28.0`, TP4 `gemma-4-31B-it`) | B70 detection, `gpu.intel.com/xe` asserted by `up`, vLLM XPU `v0.32.0` by digest through `agent up` | Host Level Zero and OpenCL libraries mounted by `hostPath` resolve inside the k3d node container, not on the host |
 | Intel GPU plugin from intel-device-plugins `main` (`intel_gpu_plugin_ref`, unpinned) | Plugin `0.37.1` baked into the node image | Two GPU plugin DaemonSets on one node |
 | NUMA-aware NRI balloons, one `BalloonsPolicy` per node | Baked balloons for hybrid P/E and multi-NUMA hosts; `K3D_GPU_BALLOONS=external` [ledger:H14] | Two policies if `deploy_nri_balloon_policy=yes`; `eat-up.sh` sets `off` |
 | Kata sandboxes (`kata-fc`, `kata-qemu`) | `doctor --perf` reports `/dev/kvm` on the host | Kata needs `/dev/kvm` inside the k3d node, which `up` does not pass through |

diff --git a/docs/platforms.md b/docs/platforms.md
@@ -323,7 +323,7 @@ B70 runs land:
   `/usr/lib/x86_64-linux-gnu/intel-opencl`).
 - What k3d-gpu already covers: B70 detection and the Intel GPU plugin `0.37.1`, baked
   and pinned; `gpu.intel.com/xe` advertised and asserted by `up`; the label through
-  `K3D_GPU_NODE_LABELS`; vLLM XPU through `K3D_GPU_AGENT_BACKEND=vllm` (`v0.31.0`, by digest).
+  `K3D_GPU_NODE_LABELS`; vLLM XPU through `K3D_GPU_AGENT_BACKEND=vllm` (`v0.32.0`, by digest).
 - Expected collision: a second GPU plugin. `deploy-intel-gpu-plugin.yml` applies
   `deployments/gpu_plugin/base/intel-gpu-plugin.yaml` from intel-device-plugins `main`
   (`intel_gpu_plugin_ref` defaults to `main`, unpinned) beside the baked DaemonSet.

diff --git a/images.lock b/images.lock
@@ -291,7 +291,7 @@
       "tag": "v0.32.0"
     },
     {
-      "digest": "sha256:95ac815038c4b3537b173798bfe3b89bdf525a52bd653cd14d4deb01f1062be7",
+      "digest": "sha256:20509fabe50dda62c09c0786d1a6effa299d3dbdddcc49331b93e65144652a8c",
       "name": "docker.io/vllm/vllm-openai-xpu",
       "sources": [
         "scripts/k3d-gpu AGENT_DEFAULT_IMAGE"
@@ -316,7 +316,7 @@
       "tag": "v0.15.0"
     },
     {
-      "digest": "sha256:be586c648cb79d2cd75d0c38a6541e1f16c47a7c6621d3e711470a467fc38744",
+      "digest": "sha256:c9aba1a673382ad356a72fb63e6cd73e8c021797e904b6c099bfe130217df62e",
       "name": "ghcr.io/ggml-org/llama.cpp",
       "sources": [
         "scripts/k3d-gpu K3D_GPU_LLAMA_VULKAN_IMAGE"

diff --git a/scripts/k3d-gpu b/scripts/k3d-gpu
@@ -7459,9 +7459,9 @@ AGENT_BACKEND="${AGENT_BACKEND,,}"
 AGENT_HPU=""    # set by agent_select when the device is a Gaudi
 AGENT_SEL=""    # "<device> <resource>", set by agent_select
 # llama.cpp server on Vulkan, the engine plan() may pick for a GPU cell other
-# than OVMS: ghcr.io/ggml-org/llama.cpp:server-vulkan-v0.5.0, the multi-arch
-# index built from release v0.5.0 (commit 7fe450e19305b828c199d602c23a8337aaa1f03b).
-LLAMA_VULKAN_IMAGE="${K3D_GPU_LLAMA_VULKAN_IMAGE:-ghcr.io/ggml-org/llama.cpp@sha256:be586c648cb79d2cd75d0c38a6541e1f16c47a7c6621d3e711470a467fc38744}"
+# than OVMS: ghcr.io/ggml-org/llama.cpp:server-vulkan-b11960, the multi-arch
+# index built from release v0.8.0 (tag b11960, commit 8642541b255bfb5dc39cd351a178be09013a55d6).
+LLAMA_VULKAN_IMAGE="${K3D_GPU_LLAMA_VULKAN_IMAGE:-ghcr.io/ggml-org/llama.cpp@sha256:c9aba1a673382ad356a72fb63e6cd73e8c021797e904b6c099bfe130217df62e}"
 # The llama.cpp agent backend: release v0.8.0 (tag b11960, commit
 # 8642541b255bfb5dc39cd351a178be09013a55d6), the latest release on 2026-10-28; its image
 # tags are server-intel-b11960 (SYCL, Level Zero and compute-runtime 26.18 inside) for an
@@ -7488,8 +7488,8 @@ agent_config() {
                 # vllm-0.30.0-ptupstream-2.13.0, pulled anonymously.
                 AGENT_DEFAULT_IMAGE="vault.habana.ai/gaudi-docker/1.24.2/ubuntu24.04/habanalabs/vllm-0.30.0-ptupstream-2.13.0@sha256:5d064465d751a73a729e7ecdc32c0ba5de0df1d305debb60a3d9a85b4fd3a967"
             elif [ "${AGENT_BACKEND}" = vllm ]; then
-                # vllm/vllm-openai-xpu:v0.31.0 (Level Zero 1.32 and compute-runtime 26.27 inside).
-                AGENT_DEFAULT_IMAGE="vllm/vllm-openai-xpu@sha256:95ac815038c4b3537b173798bfe3b89bdf525a52bd653cd14d4deb01f1062be7"
+                # vllm/vllm-openai-xpu:v0.32.0 (Level Zero 1.32 and compute-runtime 26.27 inside).
+                AGENT_DEFAULT_IMAGE="vllm/vllm-openai-xpu@sha256:20509fabe50dda62c09c0786d1a6effa299d3dbdddcc49331b93e65144652a8c"
             else
                 # intel/llm-scaler-vllm:1.5 = 0.26.0-b2 (vLLM 0.26 baseline, compute-runtime 26.14).
                 AGENT_DEFAULT_IMAGE="intel/llm-scaler-vllm@sha256:52218ad85513ab6686d4c090c83c2bd8c5b02423c63aa4dabd41837fe641fe3b"
@@ -11531,7 +11531,7 @@ agent_model_format() {
 # Applies the optimizer's choices for AGENT_OPT_MODEL to the agent settings.
 agent_optimize() {
     [ -n "${AGENT_OPT_MODEL}" ] || return 0
-    local dir bundle out errf engine=${AGENT_APP/#llamacpp/llama.cpp} device=${AGENT_SEL%% *} vver=0.31.0 draft=auto ledger meas
+    local dir bundle out errf engine=${AGENT_APP/#llamacpp/llama.cpp} device=${AGENT_SEL%% *} vver=0.32.0 draft=auto ledger meas
     local follow=() OPT_ENGINE="" OPT_MODEL="" OPT_REVISION="" OPT_TOOL_PARSER="" OPT_REASONING_PARSER="" OPT_SPEC=""
     local OPT_CONVERSION="" OPT_ARGS_JSON="[]" OPT_SUMMARY="" OPT_FALLBACK_MODEL="" OPT_FALLBACK_REVISION="" OPT_WARNING=""
     AGENT_MODEL=${AGENT_OPT_MODEL}

diff --git a/tests/bench_vs_test.sh b/tests/bench_vs_test.sh
@@ -45,7 +45,7 @@ dry_ollama() {
     local o
     o=$(dry --vs ollama --id H10 --cpuset 0-15) || return 1
     grep -q '^docker run -d --name k3dgpu-g-vs-ollama -p 127.0.0.1:[0-9]*:11434 ' <<< "${o}" \
-        && grep -q -- '--cpuset-cpus 0-15 ollama/ollama:0.40.5@sha256:5d3a2f25' <<< "${o}" \
+        && grep -q -- '--cpuset-cpus 0-15 ollama/ollama:0.40.13@sha256:b2f0122c' <<< "${o}" \
         && grep -q -- '--id H10 ' <<< "${o}" \
         && grep -q -- '--arm ollama=http://127.0.0.1:[0-9]*/v1,qwen3:4b-q4_K_M$' <<< "${o}" \
         && grep -q '^verify: sha256 /root/.ollama/models/blobs/sha256-163553aea1b1de62de7c5eb2ef5afb756b4b3133308d9ae7e42e951d8d696ef5 = 163553ae' <<< "${o}"
@@ -54,7 +54,7 @@ dry_ollama() {
 dry_llamacpp() {
     local o
     o=$(dry --vs llamacpp) || return 1
-    grep -q 'llama.cpp:server-b11706@sha256:f41d47ee.* -mu https://huggingface.co/Qwen/Qwen3-4B-GGUF/resolve/bc640142c66e1fdd12af0bd68f40445458f3869b/Qwen3-4B-Q4_K_M.gguf -m /root/.cache/llama.cpp/Qwen3-4B-Q4_K_M.gguf --alias Qwen3-4B --host 0.0.0.0 --port 8080$' <<< "${o}" \
+    grep -q 'llama.cpp:server-b11960@sha256:8052c13e.* -mu https://huggingface.co/Qwen/Qwen3-4B-GGUF/resolve/bc640142c66e1fdd12af0bd68f40445458f3869b/Qwen3-4B-Q4_K_M.gguf -m /root/.cache/llama.cpp/Qwen3-4B-Q4_K_M.gguf --alias Qwen3-4B --host 0.0.0.0 --port 8080$' <<< "${o}" \
         && ! grep -q -- ' -hf ' <<< "${o}" \
         && grep -q -- '--arm llamacpp=http://127.0.0.1:[0-9]*/v1,Qwen3-4B$' <<< "${o}" \
         && grep -q -- '--id H0 --out bench-H0-llamacpp/result.json' <<< "${o}"

diff --git a/tests/console/fallback-json.golden b/tests/console/fallback-json.golden
@@ -39,7 +39,7 @@
   "decision": "green",
   "engine": {
     "device": "Vulkan",
-    "image": "ghcr.io/ggml-org/llama.cpp@sha256:be586c648cb79d2cd75d0c38a6541e1f16c47a7c6621d3e711470a467fc38744",
+    "image": "ghcr.io/ggml-org/llama.cpp@sha256:c9aba1a673382ad356a72fb63e6cd73e8c021797e904b6c099bfe130217df62e",
     "name": "llama.cpp"
   },
   "explain": [
@@ -55,7 +55,7 @@
     "  synthetic.uhd770.ovms on gpu:0000:00:02.0: green (report synthetic-fallback-4, decode 5 tok/s); excluded: target gpu:0000:00:02.0 not usable: GuC i915/tgl_guc_70.1.1.bin 70.1.1 is older than the 70.5.1 i915 wants",
     "decision:",
     "  pick: synthetic.nvidia.llama-cpp-vulkan (decode 3 tok/s, highest of 2 green cell(s) on this host)",
-    "  engine: llama.cpp on Vulkan (ghcr.io/ggml-org/llama.cpp@sha256:be586c648cb79d2cd75d0c38a6541e1f16c47a7c6621d3e711470a467fc38744)",
+    "  engine: llama.cpp on Vulkan (ghcr.io/ggml-org/llama.cpp@sha256:c9aba1a673382ad356a72fb63e6cd73e8c021797e904b6c099bfe130217df62e)",
     "  model: Qwen/Qwen3-4B-GGUF, q4_k_m",
     "  placement: gpu:nvidia via nvidia.com/gpu",
     "  endpoint: 127.0.0.1:8080, auth loopback-key (no login, gateway key in a 0600 file)",
@@ -88,7 +88,7 @@
   "schema": "k3d-gpu.plan/v1",
   "summary": [
     "pick: synthetic.nvidia.llama-cpp-vulkan (decode 3 tok/s, highest of 2 green cell(s) on this host)",
-    "engine: llama.cpp on Vulkan (ghcr.io/ggml-org/llama.cpp@sha256:be586c648cb79d2cd75d0c38a6541e1f16c47a7c6621d3e711470a467fc38744)",
+    "engine: llama.cpp on Vulkan (ghcr.io/ggml-org/llama.cpp@sha256:c9aba1a673382ad356a72fb63e6cd73e8c021797e904b6c099bfe130217df62e)",
     "model: Qwen/Qwen3-4B-GGUF, q4_k_m",
     "placement: gpu:nvidia via nvidia.com/gpu",
     "endpoint: 127.0.0.1:8080, auth loopback-key (no login, gateway key in a 0600 file)",

diff --git a/tests/console/fallback-vulkan.golden b/tests/console/fallback-vulkan.golden
@@ -10,7 +10,7 @@ matrix: 4 cells, 4 on a target of this host
   synthetic.uhd770.ovms on gpu:0000:00:02.0: green (report synthetic-fallback-4, decode 5 tok/s); excluded: target gpu:0000:00:02.0 not usable: GuC i915/tgl_guc_70.1.1.bin 70.1.1 is older than the 70.5.1 i915 wants
 decision:
   pick: synthetic.nvidia.llama-cpp-vulkan (decode 3 tok/s, highest of 2 green cell(s) on this host)
-  engine: llama.cpp on Vulkan (ghcr.io/ggml-org/llama.cpp@sha256:be586c648cb79d2cd75d0c38a6541e1f16c47a7c6621d3e711470a467fc38744)
+  engine: llama.cpp on Vulkan (ghcr.io/ggml-org/llama.cpp@sha256:c9aba1a673382ad356a72fb63e6cd73e8c021797e904b6c099bfe130217df62e)
   model: Qwen/Qwen3-4B-GGUF, q4_k_m
   placement: gpu:nvidia via nvidia.com/gpu
   endpoint: 127.0.0.1:8080, auth loopback-key (no login, gateway key in a 0600 file)