diff --git a/.github/workflows/llama.cpp-examples.yml b/.github/workflows/llama.cpp-examples.yml
new file mode 100644
index 000000000..97cdedd41
--- /dev/null
+++ b/.github/workflows/llama.cpp-examples.yml
@@ -0,0 +1,62 @@
+# llama.cpp examples guard - project thin trigger.
+#
+# Calls the common engine .github/workflows/examples-template.yml;
+# this file only declares what varies per project: schedule,
+# concurrency, upstream repo, dispatch input forwarding, and matrix
+# parallelism. The guard loop lives in the engine. See
+# examples-guards/README.md.
+
+name: llama.cpp-examples
+
+concurrency:
+  # Three groups: schedule runs serialize; each PR serializes per number
+  # so a new push cancels the stale run; each manual run stands alone.
+  # Group names are repository-wide, so the PR group carries the project
+  # name. A bare pr-<number> would let two projects on one PR cancel
+  # each other.
+  group: ${{ github.event_name == 'schedule' && 'llama.cpp-examples-schedule' || (github.event_name == 'pull_request' && format('llama.cpp-examples-pr-{0}', github.event.pull_request.number)) || format('manual-{0}', github.run_id) }}
+  cancel-in-progress: ${{ github.event_name == 'pull_request' }}
+
+on:
+  # Daily at 03:00 Beijing time (UTC 19:00), one hour after
+  # accelerate-examples. The source repo's llama.cpp-examples cron should
+  # be turned off once this is on main, so the two guards do not both
+  # take NPUs.
+  schedule:
+    - cron: '0 19 * * *'
+  workflow_dispatch:
+    inputs:
+      target_ref:
+        description: Branch, tag, or SHA to test (empty = latest upstream release; falls back to master on API failure)
+        required: false
+        default: ''
+        type: string
+  # PR runs need no merge: GitHub executes the workflow files from the
+  # PR merge ref. paths lists only this project's files, so a PR that
+  # touches only the shared engine starts no matrix here.
+  pull_request:
+    branches: [main]
+    types: [opened, synchronize, reopened]
+    paths:
+      - 'examples-guards/llama.cpp/**'
+      - '.github/workflows/llama.cpp-examples.yml'
+
+permissions:
+  actions: read
+  contents: read
+
+jobs:
+  # Job ids cannot contain '.'. The project name stays llama.cpp; that
+  # string is what artifacts and examples-guards paths use.
+  llama-cpp-examples:
+    uses: ./.github/workflows/examples-template.yml
+    with:
+      project: llama.cpp
+      upstream_repo: ggml-org/llama.cpp
+      # Upstream's default branch is master. The engine's own default is
+      # main, which 404s when a release lookup falls back.
+      default_branch: master
+      target_ref: ${{ inputs.target_ref }}
+      # 9 supported entries, all single-card. 4 leaves room on the
+      # shared pool while the Qwen GGUF lock serializes the first download.
+      max_parallel: 4
diff --git a/examples-guards/llama.cpp/README.md b/examples-guards/llama.cpp/README.md
new file mode 100644
index 000000000..314db8f55
--- /dev/null
+++ b/examples-guards/llama.cpp/README.md
@@ -0,0 +1,48 @@
+# llama.cpp example 看护
+
+本目录是 ggml-org/llama.cpp 的 example 看护数据，给共用引擎 `.github/workflows/examples-template.yml` 用。薄触发器是 `.github/workflows/llama.cpp-examples.yml`。上游默认分支是 `master`。触发器把 `default_branch` 设成 `master`，发布查询失败时回落到这条分支。
+
+被测 ref 优先取上游最新 release。清单路径按 release `v0.6.0` 的 `examples/` 核对过，列出的目录都还在。
+
+## 清单里跑什么
+
+`supported` 只保留不改上游源码就能跑完、并且用了昇腾的条目能在日志里看到 CANN 设备行的目录。PR 运行 38228583901 上，`examples/simple` 已经这样跑通。
+
+不跑的条目：
+
+- `examples/idle`：Metal 空闲微基准
+- `examples/sycl`：要 Intel oneAPI
+- `examples/training`：全量 FP32 微调，CANN 没有对应反向算子
+- `examples/simple-chat`：上游把低于 ERROR 的日志丢掉，CANN 设备行不会出现。要看到这行就得改上游源码。只看退出码 0 会把没落到昇腾的运行当成成功
+- `examples/retrieval`：上游读循环在输入结束后不返回，喂一行会一直挂到超时。让它返回就得改 `retrieval.cpp`
+- `examples/batched`、`examples/embedding`、`examples/lookahead`、`examples/lookup`、`examples/speculative`、`examples/speculative-simple`：同一轮里 ggml-cann 在 `aclnnWeightQuantBatchMatmulV2` 报 y 不连续并中止
+- `examples/parallel`：同一轮里 `FusedInferAttentionScore` 要求非 PA 路径上的 key 连续，实际不连续并中止
+- `examples/diffusion`：运行 38249312071 里 Dream 权重已从 Hugging Face 下完并转成 Q8_0 GGUF，随后在同一个 key 不连续错误上中止
+
+每条不支持的依据写在 `examples_manifest.yaml` 里该路径的上方，带源码位置或这次 CI 作业链接，不把几条合成一段注释。
+
+`v0.6.0` 里另外有 `examples/deprecation-warning`、`examples/llama-eval`、`examples/test-cmake`。源仓清单没有它们，这次不调度。
+
+## profile
+
+`scripts/setup_example.sh` 用函数 `setup_<profile>` 解释清单里的 profile。未知 profile 在编译和下载之前退出。
+
+- `cann`：打开 `GGML_CANN`，只编 `exec` 指向的那一个 target。Qwen2.5 0.5B Instruct 的 Q4_0 GGUF 先从 huggingface.co 下，失败再从 ModelScope 下。量化选 Q4_0，因为上游 CANN 文档里 910B 的数据类型表写了 FP16、Q8_0、Q4_0、BF16。
+- `cpu`：cmake 不开 `GGML_CANN`。绿灯只表示这个目录在 aarch64 上能编完、能跑完。外部汇总不要把这些条目读成昇腾推理成功。
+- `cmake-pkg`：安装之后再用 `find_package` 编 `examples/simple-cmake-pkg`，确认安装包还能链上 CANN 后端。`LD_LIBRARY_PATH` 指到 `inst/lib`，因为消费方二进制的 RUNPATH 盖不住 `libggml` 再加载的 `libggml-cann`。
+
+镜像是 `swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.1.0-910b-ubuntu22.04-py3.12`。编译器用镜像自带的 cmake、g++、make，不改 apt 源。
+
+## 缓存
+
+下载落到 runner 的 `~/.cache`，香港池会把这棵目录留到下一次运行。workflow 不写容器 volume，也不依赖源仓的预投递。
+
+多条 `cann` 腿共用同一个 Qwen GGUF。脚本先下到本 job 的暂存文件，核对 GGUF 文件头之后再改名进缓存，并用 mkdir 锁避开并发追加写。
+
+## 设备锚点
+
+`cann`、`cmake-pkg` 的运行日志里必须出现 `CANN` 后接数字的那一行。ggml 在卡没挂上时会静默退回 CPU，进程仍然可以 0 退出。看护只在自己的运行脚本里检查这行，不改上游 example 文件。
+
+## 触发
+
+定时已经打开，每天一次，UTC 19:00，也就是北京时间 03:00，比 accelerate 的 example 看护晚一小时。源仓同项目的定时仍开着的话会和本仓抢 NPU，合入后由维护者关掉源仓那条。
diff --git a/examples-guards/llama.cpp/examples_manifest.yaml b/examples-guards/llama.cpp/examples_manifest.yaml
new file mode 100644
index 000000000..f156281ba
--- /dev/null
+++ b/examples-guards/llama.cpp/examples_manifest.yaml
@@ -0,0 +1,184 @@
+# llama.cpp example guard for the shared engine in examples-template.yml.
+# Profiles are implemented by scripts/setup_example.sh. Timeouts include a
+# cold-cache budget: this runner pool does not have the source repo's
+# pre-seeded weights. Entries that crash in ggml-cann, or that cannot
+# prove the CANN device without editing upstream sources, are unsupported.
+version: 1
+scan:
+  root: examples
+  unit: directories
+  marker: CMakeLists.txt
+  max_depth: 1
+supported:
+  - path: examples/simple
+    profile: cann
+    runner: linux-aarch64-a2-1
+    npu_devices: '0'
+    image: swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.1.0-910b-ubuntu22.04-py3.12
+    exec: build/bin/llama-simple
+    overlay_args:
+      - -m ${LLAMA_CI_MODEL}
+      - -n 16
+      - -ngl 99
+    timeout_minutes: 90
+  - path: examples/passkey
+    profile: cann
+    runner: linux-aarch64-a2-1
+    npu_devices: '0'
+    image: swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.1.0-910b-ubuntu22.04-py3.12
+    exec: build/bin/llama-passkey
+    overlay_args:
+      - -m ${LLAMA_CI_MODEL}
+      - --junk 16
+      - -ngl 99
+      - -v
+    timeout_minutes: 90
+  - path: examples/debug
+    profile: cann
+    runner: linux-aarch64-a2-1
+    npu_devices: '0'
+    image: swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.1.0-910b-ubuntu22.04-py3.12
+    exec: build/bin/llama-debug
+    overlay_args:
+      - -m ${LLAMA_CI_MODEL}
+      - -p hello
+      - -ngl 99
+      - -c 256
+      - --save-logits
+      - --logits-output-dir ${CI_OUTPUT_DIR}
+      - --no-warmup
+      - -v
+    timeout_minutes: 90
+  - path: examples/eval-callback
+    profile: cann
+    runner: linux-aarch64-a2-1
+    npu_devices: '0'
+    image: swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.1.0-910b-ubuntu22.04-py3.12
+    exec: build/bin/llama-eval-callback
+    overlay_args:
+      - -m ${LLAMA_CI_MODEL}
+      - -p hello
+      - -ngl 99
+    timeout_minutes: 90
+  - path: examples/convert-llama2c-to-ggml
+    profile: cpu
+    runner: linux-aarch64-a2-1
+    npu_devices: '0'
+    image: swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.1.0-910b-ubuntu22.04-py3.12
+    exec: build/bin/llama-convert-llama2c-to-ggml
+    overlay_args:
+      - --llama2c-model ${FIXTURE_DIR}/stories260K.bin
+      - --copy-vocab-from-model ${FIXTURE_DIR}/tok512.bin
+      - --llama2c-output-model ${CI_OUTPUT_DIR}/stories260K.gguf
+    timeout_minutes: 90
+  - path: examples/gguf
+    profile: cpu
+    runner: linux-aarch64-a2-1
+    npu_devices: '0'
+    image: swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.1.0-910b-ubuntu22.04-py3.12
+    exec: build/bin/llama-gguf
+    timeout_minutes: 90
+  - path: examples/gguf-hash
+    profile: cpu
+    runner: linux-aarch64-a2-1
+    npu_devices: '0'
+    image: swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.1.0-910b-ubuntu22.04-py3.12
+    exec: build/bin/llama-gguf-hash
+    overlay_args:
+      - ${FIXTURE_DIR}/ci-demo.gguf
+    timeout_minutes: 90
+  - path: examples/gen-docs
+    profile: cpu
+    runner: linux-aarch64-a2-1
+    npu_devices: '0'
+    image: swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.1.0-910b-ubuntu22.04-py3.12
+    exec: build/bin/llama-gen-docs
+    timeout_minutes: 90
+  - path: examples/simple-cmake-pkg
+    profile: cmake-pkg
+    runner: linux-aarch64-a2-1
+    npu_devices: '0'
+    image: swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.1.0-910b-ubuntu22.04-py3.12
+    exec: examples/simple-cmake-pkg/build/llama-simple-cmake-pkg
+    overlay_args:
+      - -m ${LLAMA_CI_MODEL}
+      - -n 16
+      - -ngl 99
+    timeout_minutes: 120
+unsupported:
+  # examples/idle/README.md links PR 17766, titled
+  # "metal : add residency sets keep-alive heartbeat". That PR's log is
+  # "ggml_metal_init: found device: Apple M2 Ultra". idle.cpp:78 sleeps
+  # to "simulate idle GPU", then times llama_process. The example measures
+  # Metal residency after idle, not a CANN workload.
+  - examples/idle
+  # ls-sycl-device.cpp:8 includes ggml-sycl.h and line 13 calls
+  # ggml_backend_sycl_print_sycl_devices(). CMakeLists.txt builds
+  # llama-ls-sycl-device. README.md says to build with GGML_SYCL_TARGET
+  # and run `source /opt/intel/oneapi/setvars.sh`. Intel oneAPI, not CANN.
+  - examples/sycl
+  # finetune.cpp:85 calls llama_opt_epoch. Lines 33-39 force an F32 KV
+  # cache because OUT_PROD has no f16 support. README.md:6-7 documents
+  # only a CPU build with no extra backend, or CUDA with -ngl 999.
+  # README.md:9 says FLASH_ATTN_EXT has no backward pass. v0.6.0
+  # ggml/src/ggml-cann/ggml-cann.cpp has no "backward" string, so the
+  # CANN backend does not implement the training backward this binary needs.
+  - examples/training
+  # simple-chat.cpp:72-77 installs llama_log_set and prints only when
+  # level >= GGML_LOG_LEVEL_ERROR. The CANN device line is below ERROR,
+  # so the guard cannot see it unless that filter is edited. Exit 0
+  # would not show that the NPU was used.
+  - examples/simple-chat
+  # retrieval.cpp:265-268 is `while (true)` then getline and
+  # common_tokenize, with no break on an empty line or EOF. One prompt
+  # does not end the process. Exiting would mean editing this loop.
+  - examples/retrieval
+  # llama-batched on v0.6.0, CANN 9.1.0, Ascend910B3 aborted in
+  # ggml-cann.cpp:70 with Invalid_Argument(EZ0026): Parameter y of
+  # aclnnWeightQuantBatchMatmulV2GetWorkspaceSize is not contiguous.
+  # Exit 134. Fixing it means editing ggml-cann, not this example.
+  # https://github.com/Ascend/docs/actions/runs/38228583901/job/114742295302
+  - examples/batched
+  # llama-embedding on run 38228583901 aborted in ggml-cann.cpp:70 with
+  # Invalid_Argument(EZ0026): Parameter y of
+  # aclnnWeightQuantBatchMatmulV2GetWorkspaceSize is not contiguous.
+  # Exit 134.
+  # https://github.com/Ascend/docs/actions/runs/38228583901/job/114742295286
+  - examples/embedding
+  # llama-lookahead on run 38228583901 aborted in ggml-cann.cpp:70 with
+  # Invalid_Argument(EZ0026): Parameter y of
+  # aclnnWeightQuantBatchMatmulV2GetWorkspaceSize is not contiguous.
+  # Exit 134.
+  # https://github.com/Ascend/docs/actions/runs/38228583901/job/114742295384
+  - examples/lookahead
+  # llama-lookup on run 38228583901 aborted in ggml-cann.cpp:70 with
+  # Invalid_Argument(EZ0026): Parameter y of
+  # aclnnWeightQuantBatchMatmulV2GetWorkspaceSize is not contiguous.
+  # Exit 134.
+  # https://github.com/Ascend/docs/actions/runs/38228583901/job/114742295213
+  - examples/lookup
+  # llama-speculative on run 38228583901 aborted in ggml-cann.cpp:70 with
+  # Invalid_Argument(EZ0026): Parameter y of
+  # aclnnWeightQuantBatchMatmulV2GetWorkspaceSize is not contiguous.
+  # Exit 134.
+  # https://github.com/Ascend/docs/actions/runs/38228583901/job/114742295204
+  - examples/speculative
+  # llama-speculative-simple on run 38228583901 aborted in ggml-cann.cpp:70
+  # with Invalid_Argument(EZ0026): Parameter y of
+  # aclnnWeightQuantBatchMatmulV2GetWorkspaceSize is not contiguous.
+  # Exit 134.
+  # https://github.com/Ascend/docs/actions/runs/38228583901/job/114742295188
+  - examples/speculative-simple
+  # llama-parallel on run 38228583901 aborted in ggml-cann.cpp:70 with
+  # Invalid_Argument(EZ0037): Parameter key of FusedInferAttentionScore
+  # is invalid because in non-PA scenarios the key tensor must be
+  # contiguous. Exit 134.
+  # https://github.com/Ascend/docs/actions/runs/38228583901/job/114742295315
+  - examples/parallel
+  # llama-diffusion-cli on run 38249312071 downloaded Dream-v0-Instruct-7B
+  # and wrote Dream-v0-Instruct-7B.Q8_0.gguf, then aborted in
+  # ggml-cann.cpp:70 with Invalid_Argument(EZ0037): Parameter key of
+  # FusedInferAttentionScore is invalid because in non-PA scenarios the
+  # key tensor must be contiguous. Exit 134.
+  # https://github.com/Ascend/docs/actions/runs/38249312071/job/114811506183
+  - examples/diffusion
diff --git a/examples-guards/llama.cpp/fixtures/ci-demo.gguf b/examples-guards/llama.cpp/fixtures/ci-demo.gguf
new file mode 100644
index 000000000..b775d9be4
Binary files /dev/null and b/examples-guards/llama.cpp/fixtures/ci-demo.gguf differ
diff --git a/examples-guards/llama.cpp/fixtures/stories260K.bin b/examples-guards/llama.cpp/fixtures/stories260K.bin
new file mode 100644
index 000000000..cfebf5a59
Binary files /dev/null and b/examples-guards/llama.cpp/fixtures/stories260K.bin differ
diff --git a/examples-guards/llama.cpp/fixtures/tok512.bin b/examples-guards/llama.cpp/fixtures/tok512.bin
new file mode 100644
index 000000000..be175e5ee
Binary files /dev/null and b/examples-guards/llama.cpp/fixtures/tok512.bin differ
diff --git a/examples-guards/llama.cpp/scripts/expand_overlay.py b/examples-guards/llama.cpp/scripts/expand_overlay.py
new file mode 100755
index 000000000..6eb5334a6
--- /dev/null
+++ b/examples-guards/llama.cpp/scripts/expand_overlay.py
@@ -0,0 +1,42 @@
+#!/usr/bin/env python3
+"""Print OVERLAY_ARGS as a shell-quoted token list.
+
+The workflow passes manifest.overlay_args as a JSON array. Each item is
+expanded against this job's environment, so ${LLAMA_CI_MODEL} and
+${CI_OUTPUT_DIR} resolve here and nowhere else.
+"""
+from __future__ import annotations
+
+import json
+import os
+import shlex
+import sys
+
+
+def main() -> int:
+    raw = os.environ.get('OVERLAY_ARGS', '').strip()
+    if not raw or raw in ('null', '""'):
+        return 0
+    try:
+        items = json.loads(raw)
+    except json.JSONDecodeError as exc:
+        print(f'OVERLAY_ARGS is not valid JSON: {exc}', file=sys.stderr)
+        return 1
+    if items in (None, ''):
+        return 0
+    if not isinstance(items, list):
+        print(f'OVERLAY_ARGS must be a JSON array, got {type(items).__name__}',
+              file=sys.stderr)
+        return 1
+    tokens: list[str] = []
+    for item in items:
+        if not isinstance(item, str) or not item.strip():
+            print('OVERLAY_ARGS items must be non-empty strings', file=sys.stderr)
+            return 1
+        tokens.extend(shlex.split(os.path.expandvars(item), posix=True))
+    print(' '.join(shlex.quote(token) for token in tokens))
+    return 0
+
+
+if __name__ == '__main__':
+    raise SystemExit(main())
diff --git a/examples-guards/llama.cpp/scripts/fetch_gguf.py b/examples-guards/llama.cpp/scripts/fetch_gguf.py
new file mode 100755
index 000000000..8cc812dcb
--- /dev/null
+++ b/examples-guards/llama.cpp/scripts/fetch_gguf.py
@@ -0,0 +1,113 @@
+#!/usr/bin/env python3
+"""Download one GGUF into the runner cache.
+
+Hugging Face first. ModelScope only if that download fails. Bytes land in
+a private .part file and are renamed into the cache only after curl
+finishes and the GGUF header checks out, so parallel matrix legs do not
+append to one shared file.
+"""
+from __future__ import annotations
+
+import argparse
+import os
+import shutil
+import subprocess
+import sys
+from pathlib import Path
+
+from nfs_lock import download_lock, sweep_staging
+
+CACHE = Path(os.environ.get(
+    'LLAMA_CPP_CACHE', os.path.expanduser('~/.cache/llama.cpp')))
+MAGIC = b'GGUF'
+
+
+def gguf_ok(path: Path, min_bytes: int) -> bool:
+    if not path.is_file():
+        return False
+    try:
+        if path.stat().st_size < min_bytes:
+            return False
+        with path.open('rb') as handle:
+            return handle.read(4) == MAGIC
+    except OSError:
+        return False
+
+
+def curl_to(url: str, part: Path) -> None:
+    part.parent.mkdir(parents=True, exist_ok=True)
+    if part.is_file():
+        with part.open('rb') as handle:
+            if handle.read(4) != MAGIC:
+                part.unlink()
+    subprocess.run(
+        [
+            'curl', '-fL', '--retry', '5', '--retry-delay', '5',
+            '--connect-timeout', '30', '-C', '-', url, '-o', str(part),
+        ],
+        check=True,
+    )
+
+
+def publish(name: str, min_bytes: int, hf_url: str, ms_url: str) -> Path:
+    dest = CACHE / name
+    if dest.exists() and not gguf_ok(dest, min_bytes):
+        dest.unlink()
+    staging = CACHE / 'staging' / f'{name}.{os.getpid()}'
+    staging.mkdir(parents=True, exist_ok=True)
+    part = staging / name
+    try:
+        print(f'[download] {name} <- HuggingFace', flush=True)
+        try:
+            curl_to(hf_url, part)
+        except subprocess.CalledProcessError as exc:
+            print(f'[fallback] HuggingFace failed ({exc}); trying ModelScope',
+                  flush=True)
+            part.unlink(missing_ok=True)
+            try:
+                curl_to(ms_url, part)
+            except subprocess.CalledProcessError as ms_exc:
+                print(f'{name}: neither HuggingFace nor ModelScope delivered '
+                      f'the file ({ms_exc}) - download noise, not an example '
+                      'failure', file=sys.stderr)
+                raise SystemExit(1) from ms_exc
+        if not gguf_ok(part, min_bytes):
+            print(f'{name}: downloaded file is not a complete GGUF',
+                  file=sys.stderr)
+            raise SystemExit(1)
+        dest.parent.mkdir(parents=True, exist_ok=True)
+        os.replace(part, dest)
+    finally:
+        shutil.rmtree(staging, ignore_errors=True)
+    print(f'model ready: {dest}', flush=True)
+    return dest
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument('--name', required=True)
+    parser.add_argument('--min-bytes', type=int, required=True)
+    parser.add_argument('--hf-url', required=True)
+    parser.add_argument('--ms-url', required=True)
+    parser.add_argument('--env-var', default='LLAMA_CI_MODEL')
+    args = parser.parse_args()
+    os.environ['HF_ENDPOINT'] = 'https://huggingface.co'
+    sweep_staging(CACHE / 'staging')
+    dest = CACHE / args.name
+    with download_lock(CACHE / 'locks', args.name):
+        if gguf_ok(dest, args.min_bytes):
+            print(f'reusing {dest}', flush=True)
+        else:
+            dest = publish(args.name, args.min_bytes, args.hf_url, args.ms_url)
+    github_env = os.environ.get('GITHUB_ENV')
+    if not github_env:
+        print('GITHUB_ENV is required', file=sys.stderr)
+        return 1
+    with open(github_env, 'a', encoding='utf-8') as handle:
+        handle.write(f'{args.env_var}={dest}\n')
+    print(f'{args.env_var}={dest}', flush=True)
+    return 0
+
+
+if __name__ == '__main__':
+    raise SystemExit(main())
diff --git a/examples-guards/llama.cpp/scripts/nfs_lock.py b/examples-guards/llama.cpp/scripts/nfs_lock.py
new file mode 100755
index 000000000..bb21aadd0
--- /dev/null
+++ b/examples-guards/llama.cpp/scripts/nfs_lock.py
@@ -0,0 +1,100 @@
+#!/usr/bin/env python3
+"""mkdir lock for the shared runner cache.
+
+fcntl locks do not exclude jobs on different pods of this NFS volume.
+mkdir is atomic across those pods. A heartbeat refreshes the lock mtime
+so a waiter can delete a lock left behind by a killed job.
+"""
+from __future__ import annotations
+
+import os
+import shutil
+import threading
+import time
+import uuid
+from collections.abc import Iterator
+from contextlib import contextmanager
+from pathlib import Path
+
+HEARTBEAT_SECONDS = 30
+LOCK_STALE_SECONDS = 300
+LOCK_POLL_SECONDS = 10
+LOCK_WAIT_SECONDS = 900
+STAGING_STALE_SECONDS = 24 * 3600
+
+
+def heartbeat(lock: Path, stop: threading.Event) -> None:
+    while not stop.wait(HEARTBEAT_SECONDS):
+        try:
+            os.utime(lock)
+        except OSError:
+            return
+
+
+def break_if_stale(lock: Path) -> None:
+    try:
+        idle = time.time() - lock.stat().st_mtime
+    except FileNotFoundError:
+        return
+    if idle < LOCK_STALE_SECONDS:
+        return
+    print(f'[lock] {lock.name}: no heartbeat for {idle:.0f}s, breaking it',
+          flush=True)
+    stale = lock.with_name(f'{lock.name}.stale-{uuid.uuid4().hex}')
+    try:
+        os.rename(lock, stale)
+    except FileNotFoundError:
+        return
+    shutil.rmtree(stale, ignore_errors=True)
+
+
+@contextmanager
+def download_lock(lock_dir: Path, name: str) -> Iterator[None]:
+    """Hold lock_dir/name for the body.
+
+    After LOCK_WAIT_SECONDS the body runs without the lock, so a stuck
+    holder cannot consume the whole job timeout.
+    """
+    lock_dir.mkdir(parents=True, exist_ok=True)
+    lock = lock_dir / name
+    deadline = time.monotonic() + LOCK_WAIT_SECONDS
+    waiting = False
+    while True:
+        try:
+            lock.mkdir()
+            break
+        except FileExistsError:
+            pass
+        if time.monotonic() >= deadline:
+            print(f'[lock] {name}: still held after {LOCK_WAIT_SECONDS}s, '
+                  'continuing without it', flush=True)
+            yield
+            return
+        if not waiting:
+            print(f'[lock] {name}: another job is downloading, waiting',
+                  flush=True)
+            waiting = True
+        break_if_stale(lock)
+        time.sleep(LOCK_POLL_SECONDS)
+    stop = threading.Event()
+    beat = threading.Thread(target=heartbeat, args=(lock, stop), daemon=True)
+    beat.start()
+    try:
+        yield
+    finally:
+        stop.set()
+        beat.join()
+        shutil.rmtree(lock, ignore_errors=True)
+
+
+def sweep_staging(staging_root: Path) -> None:
+    """Remove staging dirs left by killed jobs."""
+    if not staging_root.is_dir():
+        return
+    cutoff = time.time() - STAGING_STALE_SECONDS
+    for entry in staging_root.iterdir():
+        try:
+            if entry.stat().st_mtime < cutoff:
+                shutil.rmtree(entry, ignore_errors=True)
+        except OSError:
+            continue
diff --git a/examples-guards/llama.cpp/scripts/run_example.sh b/examples-guards/llama.cpp/scripts/run_example.sh
new file mode 100755
index 000000000..2194b2190
--- /dev/null
+++ b/examples-guards/llama.cpp/scripts/run_example.sh
@@ -0,0 +1,105 @@
+#!/usr/bin/env bash
+# Run one llama.cpp example from the CI working copy.
+# Overlay CLI args come from OVERLAY_ARGS, a JSON array. Never
+# git add, commit, or push the working copy.
+set -euo pipefail
+
+if [[ $# -lt 1 ]]; then
+  echo "usage: $0 <example-relpath>" >&2
+  exit 2
+fi
+
+EXAMPLE_REL="$1"
+TARGET_ROOT="${TARGET_ROOT:?TARGET_ROOT is required}"
+CI_OUTPUT_DIR="${CI_OUTPUT_DIR:?CI_OUTPUT_DIR is required}"
+EXEC_REL="${EXEC:-}"
+PROFILE="${PROFILE:-}"
+HERE=$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)
+
+EXAMPLE_PATH="$TARGET_ROOT/$EXAMPLE_REL"
+
+if [[ ! -e "$EXAMPLE_PATH" ]]; then
+  echo "example not found: $EXAMPLE_PATH" >&2
+  exit 1
+fi
+if [[ -n "$EXEC_REL" ]]; then
+  EXEC_PATH="$TARGET_ROOT/$EXEC_REL"
+  if [[ ! -f "$EXEC_PATH" ]]; then
+    echo "exec not found: $EXEC_PATH" >&2
+    exit 1
+  fi
+elif [[ -d "$EXAMPLE_PATH" ]]; then
+  echo "path is a directory; set exec on the manifest entry: $EXAMPLE_REL" >&2
+  exit 1
+fi
+
+mkdir -p "$CI_OUTPUT_DIR"
+
+source /usr/local/Ascend/ascend-toolkit/set_env.sh
+export HF_ENDPOINT=https://huggingface.co
+# Weights were resolved during setup. Do not let the run step re-check
+# them against the network.
+export HF_HUB_OFFLINE=1
+export TRANSFORMERS_OFFLINE=1
+
+if ! overlay=$(python3 "$HERE/expand_overlay.py"); then
+  exit 1
+fi
+eval "EXTRA_ARGS=( $overlay )"
+
+if [[ -n "$EXEC_REL" ]]; then
+  echo "running $EXEC_PATH for $EXAMPLE_REL with ${#EXTRA_ARGS[@]} overlay args"
+else
+  echo "running $EXAMPLE_PATH with ${#EXTRA_ARGS[@]} overlay args"
+fi
+if ((${#EXTRA_ARGS[@]})); then
+  printf 'overlay arg: %q\n' "${EXTRA_ARGS[@]}"
+fi
+
+cd "$TARGET_ROOT"
+export CI_OUTPUT_DIR ASCEND_RT_VISIBLE_DEVICES="${ASCEND_RT_VISIBLE_DEVICES:-0}"
+
+assert_cann_used() {
+  local run_log="$1"
+  # cpu builds are not an NPU result.
+  if [[ "$PROFILE" == "cpu" ]]; then
+    return 0
+  fi
+  if ! grep -q 'CANN[0-9]' "$run_log"; then
+    echo "run used no CANN device; check the container card mounts" >&2
+    exit 1
+  fi
+}
+
+if [[ -n "$EXEC_REL" ]]; then
+  RUN_LOG="$CI_OUTPUT_DIR/$(basename "$EXEC_REL").log"
+  case "$EXAMPLE_REL" in
+    examples/gguf)
+      "$EXEC_PATH" "$CI_OUTPUT_DIR/ci-demo.gguf" w 2>&1 | tee "$RUN_LOG"
+      "$EXEC_PATH" "$CI_OUTPUT_DIR/ci-demo.gguf" r 2>&1 | tee -a "$RUN_LOG"
+      ;;
+    *)
+      "$EXEC_PATH" "${EXTRA_ARGS[@]}" 2>&1 | tee "$RUN_LOG"
+      ;;
+  esac
+  assert_cann_used "$RUN_LOG"
+  exit 0
+fi
+
+if command -v python3 >/dev/null 2>&1; then
+  PYTHON=python3
+else
+  PYTHON=python
+fi
+case "$EXAMPLE_PATH" in
+  *.sh)
+    bash "$EXAMPLE_PATH" "${EXTRA_ARGS[@]}"
+    ;;
+  *.py)
+    "$PYTHON" "$EXAMPLE_PATH" "${EXTRA_ARGS[@]}"
+    ;;
+  *)
+    echo "unsupported example type: $EXAMPLE_PATH" >&2
+    exit 1
+    ;;
+esac
diff --git a/examples-guards/llama.cpp/scripts/setup_example.sh b/examples-guards/llama.cpp/scripts/setup_example.sh
new file mode 100755
index 000000000..07d246642
--- /dev/null
+++ b/examples-guards/llama.cpp/scripts/setup_example.sh
@@ -0,0 +1,122 @@
+#!/usr/bin/env bash
+# Prepare the CI environment for one supported llama.cpp example.
+# $1 is the manifest profile. Unknown profiles fail before any install.
+# EXEC and EXAMPLE_PATH come from the workflow as environment variables.
+set -euo pipefail
+
+if [[ $# -lt 1 ]]; then
+  echo "usage: $0 <profile>" >&2
+  exit 2
+fi
+
+PROFILE="$1"
+EXEC_REL="${EXEC:-}"
+EXAMPLE_REL="${EXAMPLE_PATH:-}"
+HERE=$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)
+
+# This pool reaches huggingface.co directly. Pin the endpoint so a
+# leftover hf-mirror value cannot 308 a bare repo id back to the origin.
+export HF_ENDPOINT=https://huggingface.co
+
+# Published qwen2.5-0.5b-instruct-q4_0.gguf is 428730208 bytes. The floor
+# rejects a truncated download; the header check is in fetch_gguf.py.
+QWEN_MIN_BYTES=300000000
+QWEN_HF_URL=https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct-GGUF/resolve/main/qwen2.5-0.5b-instruct-q4_0.gguf
+QWEN_MS_URL=https://www.modelscope.cn/models/Qwen/Qwen2.5-0.5B-Instruct-GGUF/resolve/master/qwen2.5-0.5b-instruct-q4_0.gguf
+
+require_exec() {
+  if [[ -z "$EXEC_REL" ]]; then
+    echo "EXEC is required for profile ${PROFILE}" >&2
+    exit 1
+  fi
+}
+
+source_cann() {
+  export PATH="/usr/local/sbin:$PATH"
+  source /usr/local/Ascend/ascend-toolkit/set_env.sh
+  if command -v npu-smi >/dev/null 2>&1; then
+    npu-smi info
+  fi
+}
+
+assert_cann_lib() {
+  if ! compgen -G "$1/libggml-cann.*" >/dev/null; then
+    echo "build produced no CANN backend library under $1" >&2
+    exit 1
+  fi
+}
+
+cmake_llama() {
+  local -a extra=("$@")
+  cmake -S "$TARGET_ROOT" -B "$TARGET_ROOT/build" \
+    -DCMAKE_BUILD_TYPE=Release \
+    "${extra[@]}"
+}
+
+build_target() {
+  require_exec
+  cmake --build "$TARGET_ROOT/build" --target "$(basename "$EXEC_REL")" -j "$(nproc)"
+}
+
+fetch_qwen() {
+  python3 "$HERE/fetch_gguf.py" \
+    --name qwen2.5-0.5b-instruct-q4_0.gguf \
+    --min-bytes "$QWEN_MIN_BYTES" \
+    --hf-url "$QWEN_HF_URL" \
+    --ms-url "$QWEN_MS_URL"
+}
+
+setup_cann() {
+  source_cann
+  fetch_qwen
+  cmake_llama -DGGML_CANN=on
+  build_target
+  assert_cann_lib "$TARGET_ROOT/build/bin"
+}
+
+setup_cpu() {
+  cmake_llama
+  build_target
+}
+
+setup_cmake-pkg() {
+  source_cann
+  fetch_qwen
+  cmake -S "$TARGET_ROOT" -B "$TARGET_ROOT/build" \
+    -DCMAKE_BUILD_TYPE=Release \
+    -DGGML_CANN=on \
+    -DCMAKE_INSTALL_PREFIX="$TARGET_ROOT/inst"
+  cmake --build "$TARGET_ROOT/build" -j "$(nproc)"
+  cmake --install "$TARGET_ROOT/build" --prefix "$TARGET_ROOT/inst"
+  cmake -S "$TARGET_ROOT/examples/simple-cmake-pkg" \
+    -B "$TARGET_ROOT/examples/simple-cmake-pkg/build" \
+    -DCMAKE_PREFIX_PATH="$TARGET_ROOT/inst/lib/cmake"
+  cmake --build "$TARGET_ROOT/examples/simple-cmake-pkg/build" -j "$(nproc)"
+  if ! compgen -G "$TARGET_ROOT/inst/lib/libggml-cann.*" >/dev/null \
+    && ! compgen -G "$TARGET_ROOT/build/bin/libggml-cann.*" >/dev/null; then
+    echo "build produced no CANN backend library under inst/lib or build/bin" >&2
+    exit 1
+  fi
+  local bin="$TARGET_ROOT/examples/simple-cmake-pkg/build/llama-simple-cmake-pkg"
+  if [[ ! -f "$bin" ]]; then
+    echo "cmake-pkg binary not found: $bin" >&2
+    exit 1
+  fi
+  # RUNPATH on the consumer binary does not apply to libggml's NEEDED
+  # backends. The run step sources CANN again and prepends the toolkit.
+  echo "LD_LIBRARY_PATH=$TARGET_ROOT/inst/lib" >> "$GITHUB_ENV"
+}
+
+supported_profiles() {
+  declare -F | awk '/^declare -f setup_/ { sub(/^declare -f setup_/, ""); print }' | paste -sd' ' -
+}
+
+if ! declare -F "setup_${PROFILE}" >/dev/null 2>&1; then
+  echo "unknown profile: ${PROFILE} (supported: $(supported_profiles))" >&2
+  exit 1
+fi
+
+TARGET_ROOT="${TARGET_ROOT:?TARGET_ROOT is required}"
+GITHUB_ENV="${GITHUB_ENV:?GITHUB_ENV is required}"
+
+"setup_${PROFILE}"