Scope engine image builds and verification by release target

2026-09-30 15:08:59 +08:00
parent d74a99dc33
commit 742d93a7c5
5 changed files with 214 additions and 49 deletions
--- a/.gitea/workflows/docker-build-push.yml
+++ b/.gitea/workflows/docker-build-push.yml
@@ -3,23 +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
+          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
@@ -27,10 +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: |
@@ -58,29 +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:${{ gitea.ref_name }}-cambricon-mlu370"
-          metax_image="$image_base:${{ gitea.ref_name }}-metax-c500"
-          vastai_image="$image_base:${{ gitea.ref_name }}-vastai-va16"
-
-          docker build \
-            --build-arg BASE_IMAGE=harbor.4pd.io/dooke/vllm/vllm/vllm-openai:v0.11.0 \
-            -t "$a100_image" .
-          docker build \
-            --build-arg BASE_IMAGE=harbor.4pd.io/hardcore-tech/cambricon-mlu370-pytorch:v25.01-20260228 \
-            -t "$cambricon_image" .
-          docker build \
-            --build-arg BASE_IMAGE=harbor.4pd.io/modelhubxc/enginex-metax/vllm:0.9.1 \
-            --build-arg PYTHON_BIN=/opt/conda/bin/python3 \
-            --build-arg VLLM_BIN=/opt/conda/bin/vllm \
-            -t "$metax_image" .
-          docker build \
-            --build-arg BASE_IMAGE=harbor.4pd.io/modelhubxc/sunruoxi/enginex-vastai-vllm-vacc-fix-tokenizer:VVI-26.02 \
-            -t "$vastai_image" .
-
-          for image in "$a100_image" "$cambricon_image" "$metax_image" "$vastai_image"; do
+          echo "IMAGE_BASE=$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 "$image" . ;;
+              cambricon-mlu370)
+                docker build \
+                  --build-arg BASE_IMAGE=harbor.4pd.io/hardcore-tech/cambricon-mlu370-pytorch:v25.01-20260228 \
+                  -t "$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
@@ -89,23 +106,24 @@ jobs:
            done
            test "$success" -eq 1
          done
-          echo "A100_IMAGE=$a100_image" >> "$GITEA_ENV"
-          echo "CAMBRICON_IMAGE=$cambricon_image" >> "$GITEA_ENV"
-          echo "METAX_IMAGE=$metax_image" >> "$GITEA_ENV"
-          echo "VASTAI_IMAGE=$vastai_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" ;;
-              MetaX_c-500) image="$METAX_IMAGE" ;;
-              Vastai_va16) image="$VASTAI_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"]}))')"
+            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" \
--- a/.gitea/workflows/task_info.env
+++ b/.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 MetaX_c-500 Vastai_va16"
--- a/README.md
+++ b/README.md
@@ -4,8 +4,9 @@
 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、寒武纪 MLU370 x4/x8、沐曦 C500 和 Vastai VA16
-构建四个镜像，五个 GPU 共用 `vllm_qwen25_compat` 框架名。基础镜像均来自
+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。

@@ -52,14 +53,21 @@ docker build --build-arg BASE_IMAGE=harbor.4pd.io/modelhubxc/sunruoxi/enginex-va

 ## 用 runner 发布

-1. 源码在公开仓库 `a4913994/enginex-qwen25-compat` 的 `main`；`v0.1.0`
-   已完成 A100/寒武纪镜像构建并提交镜像验证，验证结果须单独回读。
+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. 推送新的 `v*` 标签后，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，并为五个 GPU 各发起一次镜像验证；接口业务码也会被检查。
-4. 在“我的驱动列表”确认五个镜像验证通过，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。
+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。
--- a/release_plan.py
+++ b/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()
--- a/tests/test_release_plan.py
+++ b/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()