Scope engine image builds and verification by release target

.gitea/workflows/docker-build-push.yml

@@ -3,21 +3,33 @@ name: Build and verify Qwen2.5 compatibility engines
 on:
   push:
     tags:
-      - "v*"
+      - "v*-nvidia-a100"
+      - "v*-cambricon-mlu370"
+      - "v*-metax-c500"
+      - "v*-vastai-va16"
+      - "v*-all"
 
 jobs:
   docker:
     runs-on: amd64-ubuntu-24.04
+    env:
+      RELEASE_TAG: ${{ gitea.ref_name }}
+      REPOSITORY: ${{ gitea.repository }}
+      SERVER_URL: ${{ gitea.server_url }}
+      ACTOR: ${{ gitea.actor }}
     steps:
       - name: Clone tagged source
         run: |
-          git clone "${{ gitea.server_url }}/${{ gitea.repository }}.git" .
-          git checkout "${{ gitea.ref_name }}"
+          git clone "$SERVER_URL/$REPOSITORY.git" .
+          git checkout "$RELEASE_TAG"
 
       - name: Run compatibility tests
-        run: python3 -m unittest discover -s tests -q
+        run: |
+          python3 -m unittest discover -s tests -q
+          python3 -m py_compile engine_compat.py release_plan.py
+          bash -n vllm_wrapper.sh
 
-      - name: Load image metadata
+      - name: Select release targets and load image metadata
         run: |
           set -eu
           set -a
@@ -25,10 +37,9 @@ jobs:
           set +a
           test -n "$FRAMEWORK"
           test -n "$TASK_TYPE"
-          test -n "$GPU_TYPES"
+          python3 release_plan.py "$RELEASE_TAG" --gitea-env "$GITEA_ENV"
           echo "FRAMEWORK=$FRAMEWORK" >> "$GITEA_ENV"
           echo "TASK_TYPE=$TASK_TYPE" >> "$GITEA_ENV"
-          echo "GPU_TYPES=$GPU_TYPES" >> "$GITEA_ENV"
 
       - name: Validate GPU and task metadata
         run: |
@@ -56,19 +67,37 @@ jobs:
       - name: Build and push target images
         run: |
           set -eu
-          image_name="$(printf '%s' '${{ gitea.repository }}' | tr '[:upper:]' '[:lower:]' | tr '_' '-')"
+          image_name="$(printf '%s' "$REPOSITORY" | tr '[:upper:]' '[:lower:]' | tr '_' '-')"
           image_base="$DOCKER_REGISTRY/$DOCKER_USERNAME/$image_name"
-          a100_image="$image_base:${{ gitea.ref_name }}-nvidia-a100"
-          cambricon_image="$image_base:${{ gitea.ref_name }}-cambricon-mlu370"
-
+          echo "IMAGE_BASE=$image_base" >> "$GITEA_ENV"
+          for target in $IMAGE_TARGETS; do
+            if [ "$RELEASE_TARGET" = all ]; then
+              image_tag="$RELEASE_TAG-$target"
+            else
+              image_tag="$RELEASE_TAG"
+            fi
+            image="$image_base:$image_tag"
+            case "$target" in
+              nvidia-a100)
                 docker build \
                   --build-arg BASE_IMAGE=harbor.4pd.io/dooke/vllm/vllm/vllm-openai:v0.11.0 \
-            -t "$a100_image" .
+                  -t "$image" . ;;
+              cambricon-mlu370)
                 docker build \
                   --build-arg BASE_IMAGE=harbor.4pd.io/hardcore-tech/cambricon-mlu370-pytorch:v25.01-20260228 \
-            -t "$cambricon_image" .
-
-          for image in "$a100_image" "$cambricon_image"; do
+                  -t "$image" . ;;
+              metax-c500)
+                docker build \
+                  --build-arg BASE_IMAGE=harbor.4pd.io/modelhubxc/enginex-metax/vllm:0.9.1 \
+                  --build-arg PYTHON_BIN=/opt/conda/bin/python3 \
+                  --build-arg VLLM_BIN=/opt/conda/bin/vllm \
+                  -t "$image" . ;;
+              vastai-va16)
+                docker build \
+                  --build-arg BASE_IMAGE=harbor.4pd.io/modelhubxc/sunruoxi/enginex-vastai-vllm-vacc-fix-tokenizer:VVI-26.02 \
+                  -t "$image" . ;;
+              *) echo "unknown image target: $target"; exit 1 ;;
+            esac
             success=0
             for attempt in 1 2 3; do
               if docker push "$image"; then success=1; break; fi
@@ -77,22 +106,29 @@ jobs:
             done
             test "$success" -eq 1
           done
-          echo "A100_IMAGE=$a100_image" >> "$GITEA_ENV"
-          echo "CAMBRICON_IMAGE=$cambricon_image" >> "$GITEA_ENV"
 
       - name: Notify ModelHub image verification
         run: |
           set -eu
-          for gpu in $GPU_TYPES; do
-            case "$gpu" in
-              Nvidia-A100) image="$A100_IMAGE" ;;
-              Cambricon_mlu-370-x4|Cambricon_mlu-370-x8) image="$CAMBRICON_IMAGE" ;;
-              *) echo "unknown GPU target: $gpu"; exit 1 ;;
+          for mapping in $VERIFY_PAIRS; do
+            gpu="${mapping%%:*}"
+            target="${mapping#*:}"
+            case " $IMAGE_TARGETS " in
+              *" $target "*) ;;
+              *) echo "GPU target was not built: $gpu"; exit 1 ;;
             esac
-            payload="$(GPU_TYPE="$gpu" IMAGE="$image" REPO_URL='${{ gitea.server_url }}/${{ gitea.repository }}' TAG='${{ gitea.ref_name }}' ACTOR='${{ gitea.actor }}' python3 -c 'import json, os; print(json.dumps({"framework": os.environ["FRAMEWORK"], "gpuType": os.environ["GPU_TYPE"], "imageUrl": os.environ["IMAGE"], "taskType": os.environ["TASK_TYPE"], "createBy": os.environ["ACTOR"], "repoUrl": os.environ["REPO_URL"], "tag": os.environ["TAG"]}))')"
-            curl --silent --show-error --fail-with-body --location \
+            if [ "$RELEASE_TARGET" = all ]; then
+              image_tag="$RELEASE_TAG-$target"
+            else
+              image_tag="$RELEASE_TAG"
+            fi
+            image="$IMAGE_BASE:$image_tag"
+            payload="$(GPU_TYPE="$gpu" IMAGE="$image" REPO_URL="$SERVER_URL/$REPOSITORY" TAG="$RELEASE_TAG" python3 -c 'import json, os; print(json.dumps({"framework": os.environ["FRAMEWORK"], "gpuType": os.environ["GPU_TYPE"], "imageUrl": os.environ["IMAGE"], "taskType": os.environ["TASK_TYPE"], "createBy": os.environ["ACTOR"], "repoUrl": os.environ["REPO_URL"], "tag": os.environ["TAG"]}))')"
+            response="$(curl --silent --show-error --fail-with-body --location \
               --request POST 'https://modelhub.org.cn/adminApi/image-verify' \
               --header "Xc-Token: $FIXED_TOKEN" \
               --header 'Content-Type: application/json' \
-              --data-raw "$payload"
+              --data-raw "$payload")"
+            VERIFY_RESPONSE="$response" python3 -c 'import json,os,sys; body=json.loads(os.environ["VERIFY_RESPONSE"]); sys.exit(0 if body.get("code")==0 else 1)'
+            echo "submitted image verification for $gpu"
           done

.gitea/workflows/task_info.env

@@ -1,3 +1,2 @@
 FRAMEWORK=vllm_qwen25_compat
 TASK_TYPE=text-generation
-GPU_TYPES="Nvidia-A100 Cambricon_mlu-370-x4 Cambricon_mlu-370-x8"

Dockerfile

@@ -1,12 +1,17 @@
 ARG BASE_IMAGE=harbor.4pd.io/dooke/vllm/vllm/vllm-openai:v0.11.0
 FROM ${BASE_IMAGE}
 
+ARG PYTHON_BIN=python3
+ARG VLLM_BIN=vllm
+ENV MODELHUB_COMPAT_PYTHON=${PYTHON_BIN}
+
 COPY engine_compat.py /opt/modelhub/engine_compat.py
 COPY vllm_wrapper.sh /opt/modelhub/vllm_wrapper.sh
 
 RUN set -eu; \
-    python3 -m py_compile /opt/modelhub/engine_compat.py; \
-    vllm_bin="$(command -v vllm)"; \
+    "$PYTHON_BIN" -m py_compile /opt/modelhub/engine_compat.py; \
+    if [ "$VLLM_BIN" = vllm ]; then vllm_bin="$(command -v vllm)"; else vllm_bin="$VLLM_BIN"; fi; \
+    test -x "$vllm_bin"; \
     mv "$vllm_bin" "$(dirname "$vllm_bin")/vllm_real"; \
     cp /opt/modelhub/vllm_wrapper.sh "$vllm_bin"; \
     chmod 0755 "$vllm_bin" /opt/modelhub/engine_compat.py

README.md

@@ -4,10 +4,11 @@
 的方式：标签触发 `.gitea/workflows/docker-build-push.yml`，runner 检查元数据、构建并推送镜像，
 最后调用 ModelHub 的 `image-verify` 接口注册验证。
 
-当前为 A100 参考端和寒武纪 MLU370 x4/x8 构建两个镜像，三个 GPU 共用
-`vllm_qwen25_compat` 框架名。A100 基础镜像来自示例仓库；寒武纪基础镜像来自
-平台现有 MLU370 验证任务使用的镜像。是否允许一个仓库为多个 GPU 注册同一框架，
-需要以 runner 的 `validate` 和镜像验证结果为准。
+支持 A100 参考端、寒武纪 MLU370 x4/x8、沐曦 C500 和 Vastai VA16；
+一次显式全量发布构建四个镜像，五个 GPU 共用 `vllm_qwen25_compat` 框架名。
+寒武纪 x4/x8 共用一个镜像，但仍分别进行平台验证。基础镜像均来自
+平台已有验证任务使用的版本；A100 沿用示例仓库的基础镜像。每个 GPU 的注册
+与可用性仍以平台镜像验证结果和 `frameworks` 接口回读为准。
 
 ## 修复范围
 
@@ -16,11 +17,14 @@
 - 对明确写成 `TokenizersBackend` 且带 `tokenizer.json` 的仓库，改用
   `PreTrainedTokenizerFast`。
 - 对 `model_type=qwen2`、`eos_token=<|im_end|>` 且缺少聊天模板的模型，补充
-  简单的 Qwen ChatML 模板，避免 `/v1/chat/completions` 返回 400。
+  简单的 Qwen ChatML 模板，避免 `/v1/chat/completions` 返回 400；已有独立
+  `chat_template.jinja` 或 `chat_template.json` 的模型不补写。
 
-启动时在 `/tmp` 建立新模型视图，只重写 `tokenizer_config.json`。权重、模型代码和
-其他文件以符号链接指回只读 `/model`；源模型不变。没有命中上述问题时直接调用原始
-`vllm`。修复记录写在新视图的 `.modelhub-compat-report.json`，同时输出到容器日志。
+启动时在 `/tmp` 建立新视图，只重写 `tokenizer_config.json`。若基础镜像使用
+`--tokenizer` 指向单独的临时目录（Vastai 的现有启动脚本如此），则修复实际加载的
+tokenizer 目录并保留模型仓库的独立聊天模板，权重路径仍指向原模型；否则修复模型
+视图。源模型不变。没有命中问题时直接调用原始 `vllm`。修复记录写在新视图的
+`.modelhub-compat-report.json`，同时输出到容器日志。
 
 这不能修复缺失权重、尺寸不匹配、Ascend 上持续出现的 0 分输出或平台评测超时。
 尤其是分数为 0 的参考端与目标端必须分别排查；只换一个端的镜像不能保证过审。
@@ -39,17 +43,31 @@ docker build --build-arg BASE_IMAGE=harbor.4pd.io/dooke/vllm/vllm/vllm-openai:v0
   -t qwen25-compat:a100 .
 docker build --build-arg BASE_IMAGE=harbor.4pd.io/hardcore-tech/cambricon-mlu370-pytorch:v25.01-20260228 \
   -t qwen25-compat:mlu370 .
+docker build --build-arg BASE_IMAGE=harbor.4pd.io/modelhubxc/enginex-metax/vllm:0.9.1 \
+  --build-arg PYTHON_BIN=/opt/conda/bin/python3 \
+  --build-arg VLLM_BIN=/opt/conda/bin/vllm \
+  -t qwen25-compat:metax .
+docker build --build-arg BASE_IMAGE=harbor.4pd.io/modelhubxc/sunruoxi/enginex-vastai-vllm-vacc-fix-tokenizer:VVI-26.02 \
+  -t qwen25-compat:vastai .
 ```
 
 ## 用 runner 发布
 
-1. 源码已推送到公开仓库 `a4913994/enginex-qwen25-compat` 的 `main`；目前尚未推送触发构建的 `v*` 标签。
-2. 为该仓库的 runner 配置 `DOCKER_REGISTRY`、`DOCKER_USERNAME`、
-   `DOCKER_PASSWORD`、`FIXED_TOKEN`。不要把凭证写进仓库。
-3. 推送 `v*` 标签。工作流依次跑单元测试、验证 GPU/任务元数据、构建与推送镜像、
-   为 A100 和 MLU370 x4/x8 各发起一次镜像验证。
-4. 在“我的驱动列表”确认三个镜像验证通过，并检查新框架出现在对应 GPU 的
-   `frameworks` 接口中，再用少量已知缺陷模型做对照验证。
+1. 源码在公开仓库 `a4913994/enginex-qwen25-compat` 的 `main`。推送源码不会构建镜像；
+   只有带目标后缀的版本标签才触发 runner。历史 `v0.1.0` 和 `v0.2.1` 的镜像验证
+   结果仍须单独回读。
+2. runner 使用 `DOCKER_REGISTRY`、`DOCKER_USERNAME`、`DOCKER_PASSWORD`、
+   `FIXED_TOKEN`；不要把凭证写进仓库。
+3. 选择本次确实需要更新的镜像目标，推送 `vX.Y.Z-nvidia-a100`、
+   `vX.Y.Z-cambricon-mlu370`、`vX.Y.Z-metax-c500` 或 `vX.Y.Z-vastai-va16`。
+   例如 `v0.3.0-metax-c500` 只构建一份沐曦镜像并验证沐曦 C500；寒武纪标签
+   构建一份镜像并分别验证 x4/x8。只有明确使用 `vX.Y.Z-all` 才构建四份镜像、
+   发起五项验证。普通 `vX.Y.Z` 标签不触发工作流，也不会隐式全量发布。
+4. 工作流先跑单元测试和目标 GPU 元数据校验，再构建、推送选定镜像并检查平台
+   `image-verify` 接口业务码。通过本目录的 `python release_plan.py v0.3.0-metax-c500`
+   可以在打标签前预览将构建与验证的目标。
+5. 回读平台镜像验证记录，并检查新框架出现在目标 GPU 的 `frameworks` 接口中，
+   再用少量已知缺陷模型做对照验证。`pending` 不代表镜像已经可用。
 
 只有平台确认新框架可选且 A100 参考端实际使用新镜像后，才应让 `xc-agent` 把
 列表型 `extra_special_tokens` 的模型路由到此框架。当前智能体继续预检并跳过它们。

engine_compat.py

@@ -43,10 +43,11 @@ def _key(value: Any) -> str:
 
 
 def repair_tokenizer_config(
-    model_dir: Path, tokenizer: dict[str, Any]
+    model_dir: Path, tokenizer: dict[str, Any], *, tokenizer_dir: Path | None = None
 ) -> tuple[dict[str, Any], list[str]]:
     patched = dict(tokenizer)
     repairs: list[str] = []
+    effective_tokenizer_dir = tokenizer_dir or model_dir
 
     extra = patched.get("extra_special_tokens")
     if isinstance(extra, list):
@@ -76,12 +77,20 @@ def repair_tokenizer_config(
     # Only replace the known class mismatch from existing ModelHub failures.
     # Arbitrary custom tokenizer classes may require their own Python code.
     if patched.get("tokenizer_class") == "TokenizersBackend":
-        if not (model_dir / "tokenizer.json").is_file():
+        if not (effective_tokenizer_dir / "tokenizer.json").is_file():
             raise CompatibilityError("TokenizersBackend requires tokenizer.json")
         patched["tokenizer_class"] = "PreTrainedTokenizerFast"
         repairs.append("TokenizersBackend -> PreTrainedTokenizerFast")
 
-    if not patched.get("chat_template"):
+    # Transformers can load a standalone template with higher precedence than
+    # tokenizer_config.json. Do not synthesize one when the repository already
+    # supplies a template in that format.
+    has_standalone_template = any(
+        (directory / name).is_file()
+        for directory in (effective_tokenizer_dir, model_dir)
+        for name in ("chat_template.jinja", "chat_template.json")
+    )
+    if not patched.get("chat_template") and not has_standalone_template:
         config_path = model_dir / "config.json"
         if config_path.is_file():
             model_config = _json_object(config_path)
@@ -95,15 +104,37 @@ def repair_tokenizer_config(
     return patched, repairs
 
 
-def prepare_model(model_dir: str | Path, *, work_root: str | Path | None = None) -> str | None:
-    source = Path(model_dir).resolve()
+def prepare_model(
+    model_dir: str | Path,
+    *,
+    tokenizer_dir: str | Path | None = None,
+    work_root: str | Path | None = None,
+) -> str | None:
+    model_source = Path(model_dir).resolve()
+    if not model_source.is_dir():
+        raise CompatibilityError(f"model directory is missing: {model_source}")
+    source = Path(tokenizer_dir).resolve() if tokenizer_dir is not None else model_source
     if not source.is_dir():
-        raise CompatibilityError(f"model directory is missing: {source}")
+        raise CompatibilityError(f"tokenizer directory is missing: {source}")
     config_path = source / "tokenizer_config.json"
     if not config_path.is_file():
         return None
 
-    patched, repairs = repair_tokenizer_config(source, _json_object(config_path))
+    patched, repairs = repair_tokenizer_config(
+        model_source, _json_object(config_path), tokenizer_dir=source
+    )
+    # Some vendor entrypoints copy tokenizer files to /tmp and pass that path
+    # explicitly with --tokenizer, omitting a standalone template in the copy.
+    # Keep the original template visible to the effective tokenizer directory.
+    extra_templates = [
+        name
+        for name in ("chat_template.jinja", "chat_template.json")
+        if source != model_source
+        and not (source / name).is_file()
+        and (model_source / name).is_file()
+    ]
+    if extra_templates:
+        repairs.append("preserved standalone chat template")
     if not repairs:
         return None
 
@@ -130,8 +161,17 @@ def prepare_model(model_dir: str | Path, *, work_root: str | Path | None = None)
                         raise CompatibilityError(
                             f"cannot link model item: {item.name}"
                         ) from link_exc
+        for name in extra_templates:
+            (view / name).symlink_to(model_source / name)
         (view / ".modelhub-compat-report.json").write_text(
-            json.dumps({"source": str(source), "repairs": repairs}, ensure_ascii=False)
+            json.dumps(
+                {
+                    "model_source": str(model_source),
+                    "tokenizer_source": str(source),
+                    "repairs": repairs,
+                },
+                ensure_ascii=False,
+            )
             + "\n",
             encoding="utf-8",
         )
@@ -146,11 +186,13 @@ def prepare_model(model_dir: str | Path, *, work_root: str | Path | None = None)
 
 
 def main() -> int:
-    if len(sys.argv) != 2:
-        print("usage: engine_compat.py MODEL_DIR", file=sys.stderr)
+    if len(sys.argv) not in (2, 3):
+        print("usage: engine_compat.py MODEL_DIR [TOKENIZER_DIR]", file=sys.stderr)
         return 2
     try:
-        view = prepare_model(sys.argv[1])
+        view = prepare_model(
+            sys.argv[1], tokenizer_dir=sys.argv[2] if len(sys.argv) == 3 else None
+        )
     except CompatibilityError as exc:
         print(f"model compatibility error: {exc}", file=sys.stderr)
         return 2

release_plan.py

@@ -0,0 +1,74 @@
+"""Select the image builds and GPU verifications for a scoped release tag."""
+
+from __future__ import annotations
+
+import argparse
+import json
+import re
+from dataclasses import dataclass
+from pathlib import Path
+
+
+GPU_GROUPS = {
+    "nvidia-a100": ("Nvidia-A100",),
+    "cambricon-mlu370": ("Cambricon_mlu-370-x4", "Cambricon_mlu-370-x8"),
+    "metax-c500": ("MetaX_c-500",),
+    "vastai-va16": ("Vastai_va16",),
+}
+RELEASE_TAG = re.compile(
+    r"v[0-9]+\.[0-9]+\.[0-9]+-(nvidia-a100|cambricon-mlu370|metax-c500|vastai-va16|all)\Z"
+)
+
+
+@dataclass(frozen=True)
+class ReleasePlan:
+    target: str
+    image_targets: tuple[str, ...]
+    gpu_types: tuple[str, ...]
+    verification_pairs: tuple[tuple[str, str], ...]
+
+
+def plan_for_tag(tag: str) -> ReleasePlan:
+    match = RELEASE_TAG.fullmatch(tag)
+    if match is None:
+        raise ValueError(
+            "release tag must be vX.Y.Z-<target>, where target is "
+            "nvidia-a100, cambricon-mlu370, metax-c500, vastai-va16, or all"
+        )
+    target = match.group(1)
+    image_targets = tuple(GPU_GROUPS) if target == "all" else (target,)
+    verification_pairs = tuple(
+        (gpu, group) for group in image_targets for gpu in GPU_GROUPS[group]
+    )
+    gpu_types = tuple(gpu for gpu, _ in verification_pairs)
+    return ReleasePlan(target, image_targets, gpu_types, verification_pairs)
+
+
+def write_gitea_env(plan: ReleasePlan, path: Path) -> None:
+    with path.open("a", encoding="utf-8") as env_file:
+        env_file.write(f"RELEASE_TARGET={plan.target}\n")
+        env_file.write(f"IMAGE_TARGETS={' '.join(plan.image_targets)}\n")
+        env_file.write(f"GPU_TYPES={' '.join(plan.gpu_types)}\n")
+        env_file.write(
+            "VERIFY_PAIRS="
+            + " ".join(f"{gpu}:{group}" for gpu, group in plan.verification_pairs)
+            + "\n"
+        )
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("tag", help="scoped release tag, such as v0.3.0-metax-c500")
+    parser.add_argument("--gitea-env", type=Path, help="append the plan to GITEA_ENV")
+    args = parser.parse_args()
+    try:
+        plan = plan_for_tag(args.tag)
+    except ValueError as exc:
+        parser.error(str(exc))
+    if args.gitea_env:
+        write_gitea_env(plan, args.gitea_env)
+    print(json.dumps(plan.__dict__, ensure_ascii=False))
+
+
+if __name__ == "__main__":
+    main()

tests/test_engine_compat.py

@@ -53,6 +53,14 @@ class EngineCompatTests(unittest.TestCase):
             self.assertIsNone(prepare_model(source, work_root=root))
             self.assertEqual([p.name for p in root.iterdir()], ["model"])
 
+    def test_standalone_chat_template_is_preserved(self):
+        for filename in ("chat_template.jinja", "chat_template.json"):
+            with self.subTest(filename=filename), tempfile.TemporaryDirectory() as directory:
+                root = Path(directory)
+                source = self._model(root, {"eos_token": "<|im_end|>"})
+                (source / filename).write_text("{{ messages }}", encoding="utf-8")
+                self.assertIsNone(prepare_model(source, work_root=root))
+
     def test_known_tokenizer_class_is_replaced_only_with_tokenizer_json(self):
         with tempfile.TemporaryDirectory() as directory:
             root = Path(directory)
@@ -76,6 +84,47 @@ class EngineCompatTests(unittest.TestCase):
             )
             self.assertIsNone(prepare_model(source, work_root=root))
 
+    def test_explicit_tokenizer_directory_gets_qwen_template(self):
+        with tempfile.TemporaryDirectory() as directory:
+            root = Path(directory)
+            model = self._model(root, {"eos_token": "<|im_end|>"})
+            tokenizer = root / "vendor-tokenizer"
+            tokenizer.mkdir()
+            (tokenizer / "tokenizer_config.json").write_text(
+                json.dumps(
+                    {
+                        "eos_token": "<|im_end|>",
+                        "extra_special_tokens": ["<|im_start|>"],
+                        "tokenizer_class": "TokenizersBackend",
+                    }
+                ),
+                encoding="utf-8",
+            )
+            (tokenizer / "tokenizer.json").write_text("{}", encoding="utf-8")
+            view = Path(prepare_model(model, tokenizer_dir=tokenizer, work_root=root))
+            fixed = json.loads((view / "tokenizer_config.json").read_text("utf-8"))
+            self.assertEqual(fixed["chat_template"], QWEN_CHAT_TEMPLATE)
+            self.assertEqual(fixed["additional_special_tokens"], ["<|im_start|>"])
+            self.assertEqual(fixed["tokenizer_class"], "PreTrainedTokenizerFast")
+            self.assertEqual((model / "model.safetensors").read_bytes(), b"sample-weights")
+            self.assertFalse((view / "model.safetensors").exists())
+            self.assertEqual((view / "tokenizer.json").read_text("utf-8"), "{}")
+
+    def test_explicit_tokenizer_directory_preserves_model_standalone_template(self):
+        with tempfile.TemporaryDirectory() as directory:
+            root = Path(directory)
+            model = self._model(root, {"eos_token": "<|im_end|>"})
+            (model / "chat_template.jinja").write_text("{{ messages }}", encoding="utf-8")
+            tokenizer = root / "vendor-tokenizer"
+            tokenizer.mkdir()
+            (tokenizer / "tokenizer_config.json").write_text(
+                json.dumps({"eos_token": "<|im_end|>"}), encoding="utf-8"
+            )
+            view = Path(prepare_model(model, tokenizer_dir=tokenizer, work_root=root))
+            fixed = json.loads((view / "tokenizer_config.json").read_text("utf-8"))
+            self.assertNotIn("chat_template", fixed)
+            self.assertEqual((view / "chat_template.jinja").read_text("utf-8"), "{{ messages }}")
+
 
 if __name__ == "__main__":
     unittest.main()

tests/test_release_plan.py

@@ -0,0 +1,66 @@
+import tempfile
+import unittest
+from pathlib import Path
+
+from release_plan import GPU_GROUPS, plan_for_tag, write_gitea_env
+
+
+class ReleasePlanTests(unittest.TestCase):
+    def test_one_vendor_tag_builds_only_that_image_and_gpu(self):
+        plan = plan_for_tag("v0.3.0-metax-c500")
+        self.assertEqual(plan.target, "metax-c500")
+        self.assertEqual(plan.image_targets, ("metax-c500",))
+        self.assertEqual(plan.gpu_types, ("MetaX_c-500",))
+        self.assertEqual(plan.verification_pairs, (("MetaX_c-500", "metax-c500"),))
+
+    def test_cambricon_cards_reuse_one_image(self):
+        plan = plan_for_tag("v0.3.0-cambricon-mlu370")
+        self.assertEqual(plan.image_targets, ("cambricon-mlu370",))
+        self.assertEqual(
+            plan.gpu_types,
+            ("Cambricon_mlu-370-x4", "Cambricon_mlu-370-x8"),
+        )
+        self.assertEqual(
+            plan.verification_pairs,
+            (
+                ("Cambricon_mlu-370-x4", "cambricon-mlu370"),
+                ("Cambricon_mlu-370-x8", "cambricon-mlu370"),
+            ),
+        )
+
+    def test_all_requires_explicit_suffix(self):
+        plan = plan_for_tag("v0.3.0-all")
+        self.assertEqual(plan.image_targets, tuple(GPU_GROUPS))
+        self.assertEqual(len(plan.gpu_types), 5)
+        self.assertEqual(len(set(plan.gpu_types)), 5)
+
+    def test_unscoped_or_unknown_tag_cannot_default_to_all(self):
+        for tag in (
+            "v0.3.0",
+            "v0.3.0-unknown",
+            "v0.3.0-metax-c500-extra",
+            "v０.3.0-all",
+            "latest",
+        ):
+            with self.subTest(tag=tag), self.assertRaises(ValueError):
+                plan_for_tag(tag)
+
+    def test_gitea_env_contains_only_selected_targets(self):
+        with tempfile.TemporaryDirectory() as directory:
+            path = Path(directory) / "gitea_env"
+            path.write_text("FRAMEWORK=vllm_qwen25_compat\n", encoding="utf-8")
+            write_gitea_env(plan_for_tag("v0.3.0-vastai-va16"), path)
+            self.assertEqual(
+                path.read_text(encoding="utf-8").splitlines(),
+                [
+                    "FRAMEWORK=vllm_qwen25_compat",
+                    "RELEASE_TARGET=vastai-va16",
+                    "IMAGE_TARGETS=vastai-va16",
+                    "GPU_TYPES=Vastai_va16",
+                    "VERIFY_PAIRS=Vastai_va16:vastai-va16",
+                ],
+            )
+
+
+if __name__ == "__main__":
+    unittest.main()

vllm_wrapper.sh

@@ -2,6 +2,7 @@
 set -euo pipefail
 
 vllm_real="$(dirname "$(readlink -f "$0")")/vllm_real"
+python_bin="${MODELHUB_COMPAT_PYTHON:-python3}"
 
 if [[ "${1:-}" != "serve" || -z "${2:-}" ]]; then
     exec "$vllm_real" "$@"
@@ -9,17 +10,39 @@ fi
 
 model_dir="$2"
 shift 2
-prepared="$(python3 /opt/modelhub/engine_compat.py "$model_dir")"
+args=("$@")
+tokenizer_dir=""
+for ((i=0; i<${#args[@]}; i++)); do
+    if [[ "${args[i]}" == "--tokenizer" && $((i+1)) -lt ${#args[@]} ]]; then
+        tokenizer_dir="${args[i+1]}"
+        break
+    elif [[ "${args[i]}" == --tokenizer=* ]]; then
+        tokenizer_dir="${args[i]#--tokenizer=}"
+        break
+    fi
+done
+
+if [[ -n "$tokenizer_dir" && -d "$tokenizer_dir" ]]; then
+    prepared="$("$python_bin" /opt/modelhub/engine_compat.py "$model_dir" "$tokenizer_dir")"
+else
+    prepared="$("$python_bin" /opt/modelhub/engine_compat.py "$model_dir")"
+    tokenizer_dir=""
+fi
 if [[ -z "$prepared" ]]; then
-    exec "$vllm_real" serve "$model_dir" "$@"
+    exec "$vllm_real" serve "$model_dir" "${args[@]}"
 fi
 
-args=("$@")
+effective_model="$prepared"
+if [[ -n "$tokenizer_dir" && "$(readlink -f "$tokenizer_dir")" != "$(readlink -f "$model_dir")" ]]; then
+    # A vendor entrypoint's --tokenizer points at a copied tokenizer tree.
+    # Keep model weights at the original path; redirect only that tokenizer.
+    effective_model="$model_dir"
+fi
 for ((i=0; i<${#args[@]}; i++)); do
-    if [[ "${args[i]}" == "--tokenizer" && $((i+1)) -lt ${#args[@]} && "${args[i+1]}" == "$model_dir" ]]; then
+    if [[ "${args[i]}" == "--tokenizer" && $((i+1)) -lt ${#args[@]} && "${args[i+1]}" == "$tokenizer_dir" ]]; then
         args[i+1]="$prepared"
-    elif [[ "${args[i]}" == "--tokenizer=$model_dir" ]]; then
+    elif [[ "${args[i]}" == "--tokenizer=$tokenizer_dir" && -n "$tokenizer_dir" ]]; then
         args[i]="--tokenizer=$prepared"
     fi
 done
-exec "$vllm_real" serve "$prepared" "${args[@]}"
+exec "$vllm_real" serve "$effective_model" "${args[@]}"