初始提交：天垓150兼容引擎 I0–I4/I6/I7

.gitea/workflows/build.yml

@@ -0,0 +1,56 @@
+name: Docker Build
+
+on:
+  push:
+    tags:
+      - "v*"
+
+jobs:
+  docker:
+    runs-on: amd64-ubuntu-24.04
+    steps:
+      - name: Clone repository
+        run: |
+          git clone "${{ gitea.server_url }}/${{ gitea.repository }}.git" .
+          git checkout "${{ gitea.ref_name }}"
+
+      - name: Load task info
+        run: |
+          set -a
+          . .gitea/workflows/task_info.env
+          set +a
+          echo "标签: ${{ gitea.ref_name }}"
+          echo "FRAMEWORK=${FRAMEWORK:-}"
+          echo "GPU_TYPE=${GPU_TYPE:-}"
+          echo "TASK_TYPE=${TASK_TYPE:-}"
+          if [ -z "${FRAMEWORK:-}" ] || [ "${FRAMEWORK}" = "vllm_fix_tokenizer" ]; then
+            echo "FRAMEWORK 无效。本仓使用 vllm_0_17_0_corex_4_4_0。"
+            exit 1
+          fi
+          if [ "${GPU_TYPE:-}" != "Iluvatar_bi-150" ]; then
+            echo "GPU_TYPE 与本仓锁定值不一致: ${GPU_TYPE:-空}"
+            exit 1
+          fi
+          IMAGE_NAME="$(echo "${{ gitea.repository }}" | tr '[:upper:]' '[:lower:]' | tr '_' '-')"
+          echo "IMAGE_NAME=${IMAGE_NAME}" >> "$GITEA_ENV"
+          echo "IMAGE_TAG=${{ gitea.ref_name }}" >> "$GITEA_ENV"
+
+      - name: Build image
+        run: |
+          echo "开始 docker build，标签 ${IMAGE_TAG}"
+          docker build -t "${IMAGE_NAME}:${IMAGE_TAG}" .
+          echo "docker build 完成: ${IMAGE_NAME}:${IMAGE_TAG}"
+
+      - name: Push when registry credentials exist
+        run: |
+          if [ -z "${DOCKER_REGISTRY:-}" ] || [ -z "${DOCKER_USERNAME:-}" ] || [ -z "${DOCKER_PASSWORD:-}" ]; then
+            echo "未配置 DOCKER_REGISTRY、DOCKER_USERNAME 或 DOCKER_PASSWORD。"
+            echo "本次为 build-only，不登录仓库，不推送镜像。"
+            exit 0
+          fi
+          remote="${DOCKER_REGISTRY}/${DOCKER_USERNAME}/${IMAGE_NAME}:${IMAGE_TAG}"
+          echo "推送目标: ${remote}"
+          echo "$DOCKER_PASSWORD" | docker login "$DOCKER_REGISTRY" -u "$DOCKER_USERNAME" --password-stdin
+          docker tag "${IMAGE_NAME}:${IMAGE_TAG}" "$remote"
+          docker push "$remote"
+          echo "推送完成: ${remote}"

.gitea/workflows/task_info.env

@@ -0,0 +1,3 @@
+GPU_TYPE=Iluvatar_bi-150
+FRAMEWORK=vllm_0_17_0_corex_4_4_0
+TASK_TYPE=text-generation

.gitignore

@@ -0,0 +1,8 @@
+.env
+.env.*
+*.pem
+*.key
+secrets/
+__pycache__/
+*.pyc
+.pytest_cache/

Dockerfile

@@ -0,0 +1,20 @@
+# 基线 tag 来自公开仓反推，不是官方另行发布的锁定说明。
+# 官方入口若给出不同 tag，只改 BASE_IMAGE。
+ARG BASE_IMAGE=registry.iluvatar.com.cn:10443/customer/sz/vllm0.17.0-4.4.0-x86:v4.1
+FROM ${BASE_IMAGE}
+
+ENV FRAMEWORK=vllm_0_17_0_corex_4_4_0 \
+    GPU_TYPE=Iluvatar_bi-150 \
+    TASK_TYPE=text-generation \
+    MODEL_DIR=/model \
+    ARCH_UNKNOWN_POLICY=fail \
+    APPLY_TOKENIZER_PATCH=0
+
+WORKDIR /opt/engine
+COPY scripts/ /opt/engine/scripts/
+COPY patches/ /opt/engine/patches/
+COPY task_info.env.example /opt/engine/task_info.env
+
+RUN chmod +x /opt/engine/scripts/entrypoint.sh
+
+ENTRYPOINT ["/opt/engine/scripts/entrypoint.sh"]

Makefile

@@ -0,0 +1,9 @@
+.PHONY: test dry-run
+
+export PYTHONPATH := scripts
+
+test:
+	python3 -m unittest discover -s tests -v
+
+dry-run:
+	bash scripts/entrypoint.sh --dry-run

README.md

@@ -0,0 +1,67 @@
+# 天垓 150 文本生成兼容引擎
+
+本仓是 ModelHub XC 上 Iluvatar bi-150（天垓 150）的文本生成引擎：**Gitea 仓库 + CI 构建镜像**。日常引擎贡献不走 ZIP。
+
+**不代提交。** 本仓不向 modelhub.org.cn 提交适配任务，也不调用平台审核接口。推送、打 Tag、在模盒选用引擎都由人完成，步骤见 [docs/upload-runbook.md](docs/upload-runbook.md)。
+
+积分：待平台核实。成功看审核通过，以及适配任务能否选用该引擎。
+
+## 非目标
+
+- Track B（竞赛）为 **NO-GO**。没有 `contest/`，没有竞赛构建。
+- 没有 ZIP 打包，没有上传驱动的 Makefile 目标。
+- `FRAMEWORK` 使用 `vllm_0_17_0_corex_4_4_0`，不使用 `vllm_fix_tokenizer`。
+- tokenizer 修补（I5）未开启。现有 bi150 引擎已覆盖 tokenizer；本仓补的是 chat_template（R2）和 architecture 白名单（R4）。
+
+## 基线镜像
+
+默认：
+
+`registry.iluvatar.com.cn:10443/customer/sz/vllm0.17.0-4.4.0-x86:v4.1`
+
+这个 tag 来自公开仓反推，置信度为中高，不是官方另行发布的锁定说明。官方入口若给出不同 tag，只改 [Dockerfile](Dockerfile) 里的 `BASE_IMAGE`。
+
+构建覆盖：
+
+```bash
+docker build --build-arg BASE_IMAGE=<官方 tag> -t modelhub-engine-bi150-compat:dev .
+```
+
+无卡环境只能验证修补逻辑，不能证明吞吐。
+
+## 本地检查
+
+需要本机 `python3`、`bash`、`make`。不需要 GPU，也不需要平台账号。
+
+```bash
+make test
+make dry-run
+```
+
+`make dry-run` 只打印启动计划与模板预览，不启动 vLLM，不访问平台。
+
+## 改 FRAMEWORK
+
+默认值写在 [task_info.env.example](task_info.env.example)。镜像内和 CI 使用同一组值：
+
+- `GPU_TYPE=Iluvatar_bi-150`
+- `FRAMEWORK=vllm_0_17_0_corex_4_4_0`
+- `TASK_TYPE=text-generation`
+
+若平台给出新的 FRAMEWORK 名，同时改这三处，并让 `make test` 保持通过：
+
+- `task_info.env.example`
+- `.gitea/workflows/task_info.env`
+- `Dockerfile` 的 `ENV`
+
+不要改回 `vllm_fix_tokenizer`。
+
+## 建仓与选用
+
+见 [docs/upload-runbook.md](docs/upload-runbook.md)。事项状态见 [docs/issues.md](docs/issues.md)。证据模板见 [docs/evidence.md](docs/evidence.md)。
+
+参考公开仓的启动包装和架构改写思路，没有整仓复制：
+
+- zhengzhongwei/vLLM-Iluvatar_bi-100-opt
+- wanglin/enginex-bi100-compat
+- 公开 bi150 tokenizer 引擎只覆盖 tokenizer，本仓补 R2/R4

docs/evidence.md

@@ -0,0 +1,14 @@
+# bi150-compat 证据
+
+时区：**Asia/Shanghai**。本模板由人工填写。仓库不自动采集审核状态。
+
+- 镜像 tag：待填写
+- CI 地址：待填写
+- 审核：未开始
+- 审核通过时间（Asia/Shanghai）：待填写
+- 首次被适配任务选用的任务 ID：待填写
+- 选用时间（Asia/Shanghai）：待填写
+- 平台状态：未推送
+- 积分：待平台核实
+
+无卡本地测试只说明 R2/R4 修补逻辑能跑，不说明天垓 150 上的吞吐。

docs/issues.md

@@ -0,0 +1,14 @@
+# bi150-compat 事项
+
+Track B 不在本仓实现，状态保持 NO-GO。
+
+- [x] I0 仓骨架 + README + 非目标声明
+- [x] I1 Dockerfile + task_info.env.example
+- [x] I2 wrapper + dry-run
+- [x] I3 R2 chat_template 修补
+- [x] I4 R4 architecture 白名单
+- [ ] I5 tokenizer 修补（未开启；等 R2/R4 确认盖不住目标模型族再做）
+- [x] I6 CI build（Tag 触发；无凭据时 build-only；不提交平台任务）
+- [x] I7 人工 runbook + evidence 模板
+
+本地 `make test` 已覆盖 I3/I4 与 dry-run。镜像是否能在 Runner 上拉到基线 tag，要等第一次 `v*` 流水线。

docs/upload-runbook.md

@@ -0,0 +1,39 @@
+# 建仓、打 Tag、选用引擎（人工）
+
+本仓没有自动提交命令。CI 只构建镜像；仓库凭据齐全时才推送。CI 不调用 modelhub.org.cn，不代领适配任务。
+
+积分：待平台核实。
+
+## 1. 在 Gitea 建仓
+
+1. 在模盒的 Gitea 上新建仓库，名称建议 `modelhub-engine-bi150-compat`。
+2. 把本目录推上去。不要提交 `.env`、密钥或 Token。
+3. 确认默认分支上有 `Dockerfile`、`.gitea/workflows/build.yml` 和 `.gitea/workflows/task_info.env`。
+
+## 2. 打 Tag
+
+CI 在 `v*` 标签上触发，例如 `v0.1.0`。
+
+```bash
+git tag v0.1.0
+git push origin v0.1.0
+```
+
+标签名会成为镜像标签。
+
+## 3. 看 CI
+
+1. 打开该 Tag 的 Actions / 流水线。
+2. `Load task info` 应打印 `FRAMEWORK=vllm_0_17_0_corex_4_4_0` 和 `GPU_TYPE=Iluvatar_bi-150`。
+3. `Build image` 执行 `docker build`。基线镜像拉取失败时，日志会停在构建步骤；无卡不能代替真机验证。
+4. 若 Runner 没有 `DOCKER_REGISTRY`、`DOCKER_USERNAME`、`DOCKER_PASSWORD`，流水线打印 build-only 并成功结束，不推送。
+5. 三份凭据都由 Runner 环境提供，不写进仓库。有凭据时才会 `docker login` 并推送。
+
+## 4. 在模盒选用引擎
+
+1. 用浏览器登录模盒。
+2. 等镜像审核通过。审核状态由人在页面上确认。
+3. 在适配任务里选用该引擎。本仓不代为创建或提交任务。
+4. 把审核通过时间、镜像 tag、以及首次被选用的适配任务 ID 写入 `docs/evidence.md`。积分栏保持「待平台核实」。
+
+字段以登录后的控制台为准。

docs/weekly-report.md

@@ -0,0 +1,7 @@
+【引擎 bi150-compat】2026-10-13
+- 本周：I0–I4、I6、I7 完成；I5 未开启。阻塞：尚未推送 Gitea，也还没有真机流水线。
+- 镜像 tag：未构建。默认基线 registry.iluvatar.com.cn:10443/customer/sz/vllm0.17.0-4.4.0-x86:v4.1（公开仓反推）
+- 本地 test：见 make test
+- 平台：未推送
+- 积分：待核实
+- 下周：人工建仓、打 v* Tag、看 CI；审核通过或被选用后再填 docs/evidence.md

patches/architectures/aliases.json

@@ -0,0 +1 @@
+{}

patches/architectures/whitelist.txt

@@ -0,0 +1,13 @@
+# 本仓本地兼容白名单，不是平台公布的完整列表。
+# 只收录 vLLM 0.17 文本生成里常见、且本引擎准备接受的架构名。
+LlamaForCausalLM
+Qwen2ForCausalLM
+Qwen3ForCausalLM
+Qwen3MoeForCausalLM
+MistralForCausalLM
+MixtralForCausalLM
+Gemma2ForCausalLM
+Phi3ForCausalLM
+InternLM2ForCausalLM
+BaichuanForCausalLM
+ChatGLMForConditionalGeneration

patches/chat_template/default_chat_template.jinja

@@ -0,0 +1 @@
+{% for message in messages %}{{ '<|' + message['role'] + '|>\n' + message['content'] + '<|end|>\n' }}{% endfor %}{% if add_generation_prompt %}{{ '<|assistant|>\n' }}{% endif %}

patches/tokenizer/README.md

@@ -0,0 +1,5 @@
+# tokenizer 修补（I5，未开启）
+
+现有 bi150 引擎已经覆盖 tokenizer。本仓先补 chat_template（R2）和 architecture 白名单（R4）。
+
+`APPLY_TOKENIZER_PATCH` 默认为 `0`。在确认某类模型只靠 R2/R4 仍无法加载之前，不要打开这项。

scripts/apply_patches.py

@@ -0,0 +1,187 @@
+#!/usr/bin/env python3
+"""把 R2 chat_template 与 R4 architecture 白名单应用到模型目录。
+
+不访问平台，不启动 vLLM。
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import sys
+from pathlib import Path
+
+FORBIDDEN_FRAMEWORK = "vllm_fix_tokenizer"
+TOKENIZER_PATCH_KEY = "modelhub_tokenizer_patch"
+
+
+class PatchError(Exception):
+    """模型目录不满足本引擎的修补前提。"""
+
+
+def engine_root() -> Path:
+    return Path(__file__).resolve().parents[1]
+
+
+def assert_framework(framework: str) -> None:
+    if not framework or framework == FORBIDDEN_FRAMEWORK:
+        raise PatchError(
+            "FRAMEWORK 必须是 vllm_0_17_0_corex_4_4_0，"
+            "不能使用 vllm_fix_tokenizer"
+        )
+
+
+def load_whitelist(path: Path) -> set[str]:
+    names: set[str] = set()
+    for line in path.read_text(encoding="utf-8").splitlines():
+        item = line.split("#", 1)[0].strip()
+        if item:
+            names.add(item)
+    if not names:
+        raise PatchError(f"架构白名单为空: {path}")
+    return names
+
+
+def load_aliases(path: Path) -> dict[str, str]:
+    data = json.loads(path.read_text(encoding="utf-8"))
+    if not isinstance(data, dict):
+        raise PatchError(f"别名文件必须是对象: {path}")
+    return {str(key): str(value) for key, value in data.items()}
+
+
+def preview_default_template(messages: list[dict[str, str]], add_generation_prompt: bool = True) -> str:
+    """与 default_chat_template.jinja 同一套角色标记，供无 Jinja 依赖的本地预览。"""
+    parts: list[str] = []
+    for message in messages:
+        role = message["role"]
+        content = message["content"]
+        parts.append(f"<|{role}|>\n{content}<|end|>\n")
+    if add_generation_prompt:
+        parts.append("<|assistant|>\n")
+    return "".join(parts)
+
+
+def apply_chat_template(model_dir: Path, template_text: str, force: bool = False) -> dict:
+    path = model_dir / "tokenizer_config.json"
+    if not path.is_file():
+        raise PatchError(f"缺少 tokenizer_config.json: {path}")
+    data = json.loads(path.read_text(encoding="utf-8"))
+    if not isinstance(data, dict):
+        raise PatchError("tokenizer_config.json 必须是对象")
+    current = data.get("chat_template")
+    if isinstance(current, str) and current.strip() and not force:
+        return {"injected": False, "reason": "已有 chat_template"}
+    data["chat_template"] = template_text
+    path.write_text(json.dumps(data, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
+    return {"injected": True, "reason": "已写入 chat_template"}
+
+
+def apply_architectures(
+    model_dir: Path,
+    whitelist: set[str],
+    aliases: dict[str, str],
+    unknown_policy: str = "fail",
+) -> dict:
+    if unknown_policy not in {"fail", "skip"}:
+        raise PatchError("ARCH_UNKNOWN_POLICY 只能是 fail 或 skip")
+    path = model_dir / "config.json"
+    if not path.is_file():
+        raise PatchError(f"缺少 config.json: {path}")
+    data = json.loads(path.read_text(encoding="utf-8"))
+    if not isinstance(data, dict):
+        raise PatchError("config.json 必须是对象")
+    architectures = data.get("architectures")
+    if not isinstance(architectures, list) or not architectures:
+        raise PatchError("config.json 缺少 architectures")
+
+    rewritten: list[str] = []
+    skipped: list[str] = []
+    updated: list[str] = []
+    for name in architectures:
+        if not isinstance(name, str) or not name:
+            raise PatchError("architectures 含有空名称")
+        if name in whitelist:
+            updated.append(name)
+            continue
+        target = aliases.get(name)
+        if target in whitelist:
+            updated.append(target)
+            rewritten.append(f"{name}->{target}")
+            continue
+        if unknown_policy == "skip":
+            skipped.append(name)
+            updated.append(name)
+            continue
+        raise PatchError(f"未知架构: {name}")
+
+    changed = updated != architectures
+    if changed:
+        data["architectures"] = updated
+        path.write_text(json.dumps(data, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
+    return {
+        "architectures": updated,
+        "rewritten": rewritten,
+        "skipped": skipped,
+        "changed": changed,
+    }
+
+
+def apply_tokenizer_patch(model_dir: Path, enabled: bool) -> dict:
+    if not enabled:
+        return {"applied": False}
+    path = model_dir / "tokenizer_config.json"
+    if not path.is_file():
+        raise PatchError(f"缺少 tokenizer_config.json: {path}")
+    data = json.loads(path.read_text(encoding="utf-8"))
+    data[TOKENIZER_PATCH_KEY] = "optional-v1"
+    path.write_text(json.dumps(data, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
+    return {"applied": True}
+
+
+def apply_all(
+    model_dir: Path,
+    template_text: str,
+    whitelist: set[str],
+    aliases: dict[str, str],
+    unknown_policy: str = "fail",
+    force_chat_template: bool = False,
+    tokenizer_patch: bool = False,
+) -> dict:
+    chat = apply_chat_template(model_dir, template_text, force=force_chat_template)
+    architectures = apply_architectures(model_dir, whitelist, aliases, unknown_policy)
+    tokenizer = apply_tokenizer_patch(model_dir, tokenizer_patch)
+    return {"chat_template": chat, "architectures": architectures, "tokenizer": tokenizer}
+
+
+def main(argv: list[str] | None = None) -> int:
+    root = engine_root()
+    parser = argparse.ArgumentParser(description="应用 R2/R4 模型目录修补")
+    parser.add_argument("--model-dir", required=True, type=Path)
+    parser.add_argument("--framework", default="vllm_0_17_0_corex_4_4_0")
+    parser.add_argument("--unknown-policy", default="fail", choices=("fail", "skip"))
+    parser.add_argument("--force-chat-template", action="store_true")
+    parser.add_argument("--tokenizer-patch", action="store_true")
+    parser.add_argument("--template", type=Path, default=root / "patches/chat_template/default_chat_template.jinja")
+    parser.add_argument("--whitelist", type=Path, default=root / "patches/architectures/whitelist.txt")
+    parser.add_argument("--aliases", type=Path, default=root / "patches/architectures/aliases.json")
+    args = parser.parse_args(argv)
+    try:
+        assert_framework(args.framework)
+        result = apply_all(
+            args.model_dir,
+            args.template.read_text(encoding="utf-8"),
+            load_whitelist(args.whitelist),
+            load_aliases(args.aliases),
+            unknown_policy=args.unknown_policy,
+            force_chat_template=args.force_chat_template,
+            tokenizer_patch=args.tokenizer_patch,
+        )
+    except (PatchError, OSError, json.JSONDecodeError) as exc:
+        print(f"修补失败: {exc}", file=sys.stderr)
+        return 1
+    print(json.dumps(result, ensure_ascii=False))
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

scripts/entrypoint.sh

@@ -0,0 +1,125 @@
+#!/usr/bin/env bash
+# 读取 task_info.env，修补模型目录，并准备 vLLM 启动命令。
+# --dry-run 只打印计划，不启动进程，不访问平台。
+set -euo pipefail
+
+ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+
+usage() {
+  echo "用法: entrypoint.sh [--dry-run]" >&2
+}
+
+DRY_RUN=0
+if [[ "${1:-}" == "--dry-run" ]]; then
+  DRY_RUN=1
+elif [[ "${1:-}" == "-h" || "${1:-}" == "--help" ]]; then
+  usage
+  exit 0
+elif [[ -n "${1:-}" ]]; then
+  usage
+  exit 1
+fi
+
+info_file="${TASK_INFO_FILE:-}"
+if [[ -z "$info_file" ]]; then
+  if [[ -f "${ROOT}/task_info.env" ]]; then
+    info_file="${ROOT}/task_info.env"
+  else
+    info_file="${ROOT}/task_info.env.example"
+  fi
+fi
+
+if [[ ! -f "$info_file" ]]; then
+  echo "找不到任务信息文件: ${info_file}" >&2
+  exit 1
+fi
+
+# 已在环境中的变量优先于文件，方便本地覆盖。
+while IFS= read -r line || [[ -n "$line" ]]; do
+  case "$line" in
+    ""|\#*) continue ;;
+  esac
+  key="${line%%=*}"
+  value="${line#*=}"
+  if [[ -z "$key" || "$key" == "$line" ]]; then
+    echo "任务信息行无法解析: ${line}" >&2
+    exit 1
+  fi
+  if [[ -z "${!key+x}" ]]; then
+    export "${key}=${value}"
+  fi
+done <"$info_file"
+
+FRAMEWORK="${FRAMEWORK:-}"
+GPU_TYPE="${GPU_TYPE:-}"
+TASK_TYPE="${TASK_TYPE:-text-generation}"
+MODEL_DIR="${MODEL_DIR:-/model}"
+SERVED_MODEL_NAME="${SERVED_MODEL_NAME:-llm}"
+PORT="${PORT:-8000}"
+ARCH_UNKNOWN_POLICY="${ARCH_UNKNOWN_POLICY:-fail}"
+APPLY_TOKENIZER_PATCH="${APPLY_TOKENIZER_PATCH:-0}"
+FORCE_CHAT_TEMPLATE="${FORCE_CHAT_TEMPLATE:-0}"
+
+if [[ -z "$FRAMEWORK" || "$FRAMEWORK" == "vllm_fix_tokenizer" ]]; then
+  echo "拒绝 FRAMEWORK=${FRAMEWORK:-空}。本仓使用 vllm_0_17_0_corex_4_4_0。" >&2
+  exit 1
+fi
+if [[ -z "$GPU_TYPE" ]]; then
+  echo "GPU_TYPE 为空" >&2
+  exit 1
+fi
+
+plan=(
+  python3 -m vllm.entrypoints.openai.api_server
+  --model "$MODEL_DIR"
+  --port "$PORT"
+  --served-model-name "$SERVED_MODEL_NAME"
+  --trust-remote-code
+)
+
+echo "FRAMEWORK=${FRAMEWORK}"
+echo "GPU_TYPE=${GPU_TYPE}"
+echo "TASK_TYPE=${TASK_TYPE}"
+echo "MODEL_DIR=${MODEL_DIR}"
+echo "ARCH_UNKNOWN_POLICY=${ARCH_UNKNOWN_POLICY}"
+echo "plan: ${plan[*]}"
+
+if [[ "$DRY_RUN" -eq 1 ]]; then
+  preview="$(
+    python3 - "$ROOT" <<'PY'
+import sys
+from pathlib import Path
+root = Path(sys.argv[1])
+sys.path.insert(0, str(root / "scripts"))
+import apply_patches
+messages = [
+    {"role": "system", "content": "你是助手"},
+    {"role": "user", "content": "你好"},
+]
+print(apply_patches.preview_default_template(messages).replace("\n", "\\n"))
+PY
+  )"
+  echo "template_preview: ${preview}"
+  echo "dry-run: 不启动进程，不访问平台"
+  exit 0
+fi
+
+if [[ ! -d "$MODEL_DIR" ]]; then
+  echo "模型目录不存在: ${MODEL_DIR}" >&2
+  exit 1
+fi
+
+patch_args=(
+  --model-dir "$MODEL_DIR"
+  --framework "$FRAMEWORK"
+  --unknown-policy "$ARCH_UNKNOWN_POLICY"
+)
+if [[ "$FORCE_CHAT_TEMPLATE" == "1" ]]; then
+  patch_args+=(--force-chat-template)
+fi
+if [[ "$APPLY_TOKENIZER_PATCH" == "1" ]]; then
+  patch_args+=(--tokenizer-patch)
+fi
+
+python3 "${ROOT}/scripts/apply_patches.py" "${patch_args[@]}"
+exec "${plan[@]}"

task_info.env.example

@@ -0,0 +1,12 @@
+# 本地与镜像内的默认任务信息。平台若给出不同 FRAMEWORK，只改这一处，并保持 Dockerfile 的 ENV 与 CI 文件一致。
+GPU_TYPE=Iluvatar_bi-150
+FRAMEWORK=vllm_0_17_0_corex_4_4_0
+TASK_TYPE=text-generation
+MODEL_DIR=/model
+SERVED_MODEL_NAME=llm
+PORT=8000
+# 未知架构：fail 拒绝启动；skip 只记录并继续。默认拒绝。
+ARCH_UNKNOWN_POLICY=fail
+# I5 tokenizer 修补默认关闭。
+APPLY_TOKENIZER_PATCH=0
+FORCE_CHAT_TEMPLATE=0

tests/fixtures/fake-model/config.json

@@ -0,0 +1,4 @@
+{
+  "model_type": "llama",
+  "architectures": ["LlamaForCausalLM"]
+}

tests/fixtures/fake-model/tokenizer.json

@@ -0,0 +1,3 @@
+{
+  "version": "1.0"
+}

tests/fixtures/fake-model/tokenizer_config.json

@@ -0,0 +1,3 @@
+{
+  "tokenizer_class": "LlamaTokenizer"
+}

tests/test_contract.py

@@ -0,0 +1,96 @@
+import os
+import subprocess
+import unittest
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parents[1]
+
+
+class EngineContractTest(unittest.TestCase):
+    def test_locked_values_match(self) -> None:
+        example = (ROOT / "task_info.env.example").read_text(encoding="utf-8")
+        workflow_env = (ROOT / ".gitea/workflows/task_info.env").read_text(encoding="utf-8")
+        dockerfile = (ROOT / "Dockerfile").read_text(encoding="utf-8")
+        self.assertIn("FRAMEWORK=vllm_0_17_0_corex_4_4_0", example)
+        self.assertIn("GPU_TYPE=Iluvatar_bi-150", example)
+        self.assertIn("FRAMEWORK=vllm_0_17_0_corex_4_4_0", workflow_env)
+        self.assertIn("GPU_TYPE=Iluvatar_bi-150", workflow_env)
+        self.assertIn("FRAMEWORK=vllm_0_17_0_corex_4_4_0", dockerfile)
+        self.assertIn("GPU_TYPE=Iluvatar_bi-150", dockerfile)
+        self.assertIn(
+            "registry.iluvatar.com.cn:10443/customer/sz/vllm0.17.0-4.4.0-x86:v4.1",
+            dockerfile,
+        )
+        self.assertNotIn("FRAMEWORK=vllm_fix_tokenizer", example)
+        self.assertNotIn("FRAMEWORK=vllm_fix_tokenizer", workflow_env)
+        self.assertNotIn("FRAMEWORK=vllm_fix_tokenizer", dockerfile)
+
+    def test_ci_does_not_submit_platform_task(self) -> None:
+        workflow = (ROOT / ".gitea/workflows/build.yml").read_text(encoding="utf-8")
+        self.assertNotIn("adminApi", workflow)
+        self.assertNotIn("modelhub.org.cn", workflow)
+        self.assertIn("build-only", workflow)
+        self.assertIn('tags:', workflow)
+
+    def test_no_zip_or_contest_targets(self) -> None:
+        makefile = (ROOT / "Makefile").read_text(encoding="utf-8")
+        self.assertNotIn("package", makefile)
+        self.assertNotIn("submit", makefile)
+        self.assertFalse((ROOT / "contest").exists())
+        self.assertFalse((ROOT / "computility-run.yaml").exists())
+
+    def test_dry_run_prints_plan(self) -> None:
+        env = os.environ.copy()
+        env.pop("FRAMEWORK", None)
+        env.pop("GPU_TYPE", None)
+        env.pop("MODEL_DIR", None)
+        completed = subprocess.run(
+            ["bash", "scripts/entrypoint.sh", "--dry-run"],
+            cwd=ROOT,
+            env=env,
+            text=True,
+            capture_output=True,
+            check=False,
+        )
+        self.assertEqual(completed.returncode, 0, completed.stderr)
+        self.assertIn("FRAMEWORK=vllm_0_17_0_corex_4_4_0", completed.stdout)
+        self.assertIn("GPU_TYPE=Iluvatar_bi-150", completed.stdout)
+        self.assertIn("python3 -m vllm.entrypoints.openai.api_server", completed.stdout)
+        self.assertIn("dry-run: 不启动进程，不访问平台", completed.stdout)
+        self.assertNotIn("modelhub.org.cn", completed.stdout)
+        self.assertIn("<|user|>", completed.stdout)
+
+    def test_dry_run_rejects_forbidden_framework(self) -> None:
+        env = os.environ.copy()
+        env["FRAMEWORK"] = "vllm_fix_tokenizer"
+        completed = subprocess.run(
+            ["bash", "scripts/entrypoint.sh", "--dry-run"],
+            cwd=ROOT,
+            env=env,
+            text=True,
+            capture_output=True,
+            check=False,
+        )
+        self.assertNotEqual(completed.returncode, 0)
+        self.assertIn("vllm_fix_tokenizer", completed.stderr)
+
+    def test_tokenizer_patch_stays_off(self) -> None:
+        import shutil
+        import tempfile
+
+        import apply_patches
+
+        with tempfile.TemporaryDirectory() as tmp:
+            model = Path(tmp) / "model"
+            shutil.copytree(ROOT / "tests" / "fixtures" / "fake-model", model)
+            result = apply_patches.apply_all(
+                model,
+                (ROOT / "patches/chat_template/default_chat_template.jinja").read_text(encoding="utf-8"),
+                apply_patches.load_whitelist(ROOT / "patches/architectures/whitelist.txt"),
+                {},
+            )
+        self.assertFalse(result["tokenizer"]["applied"])
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/test_r2_chat_template.py

@@ -0,0 +1,67 @@
+import json
+import shutil
+import tempfile
+import unittest
+from pathlib import Path
+
+import apply_patches
+
+ROOT = Path(__file__).resolve().parents[1]
+FIXTURE = ROOT / "tests" / "fixtures" / "fake-model"
+TEMPLATE = (ROOT / "patches" / "chat_template" / "default_chat_template.jinja").read_text(encoding="utf-8")
+WHITELIST = apply_patches.load_whitelist(ROOT / "patches" / "architectures" / "whitelist.txt")
+ALIASES = apply_patches.load_aliases(ROOT / "patches" / "architectures" / "aliases.json")
+
+
+class ChatTemplateTest(unittest.TestCase):
+    def setUp(self) -> None:
+        self.tmp = tempfile.TemporaryDirectory()
+        self.model = Path(self.tmp.name) / "model"
+        shutil.copytree(FIXTURE, self.model)
+
+    def tearDown(self) -> None:
+        self.tmp.cleanup()
+
+    def test_injects_missing_template(self) -> None:
+        result = apply_patches.apply_chat_template(self.model, TEMPLATE)
+        self.assertTrue(result["injected"])
+        data = json.loads((self.model / "tokenizer_config.json").read_text(encoding="utf-8"))
+        self.assertEqual(data["chat_template"], TEMPLATE)
+
+    def test_keeps_existing_template(self) -> None:
+        path = self.model / "tokenizer_config.json"
+        data = json.loads(path.read_text(encoding="utf-8"))
+        data["chat_template"] = "already-set"
+        path.write_text(json.dumps(data), encoding="utf-8")
+        result = apply_patches.apply_chat_template(self.model, TEMPLATE)
+        self.assertFalse(result["injected"])
+        kept = json.loads(path.read_text(encoding="utf-8"))
+        self.assertEqual(kept["chat_template"], "already-set")
+
+    def test_force_overwrites_template(self) -> None:
+        path = self.model / "tokenizer_config.json"
+        data = json.loads(path.read_text(encoding="utf-8"))
+        data["chat_template"] = "already-set"
+        path.write_text(json.dumps(data), encoding="utf-8")
+        result = apply_patches.apply_chat_template(self.model, TEMPLATE, force=True)
+        self.assertTrue(result["injected"])
+
+    def test_missing_tokenizer_config_fails(self) -> None:
+        (self.model / "tokenizer_config.json").unlink()
+        with self.assertRaises(apply_patches.PatchError):
+            apply_patches.apply_chat_template(self.model, TEMPLATE)
+
+    def test_preview_renders_roles(self) -> None:
+        text = apply_patches.preview_default_template(
+            [
+                {"role": "system", "content": "你是助手"},
+                {"role": "user", "content": "你好"},
+            ]
+        )
+        self.assertIn("<|user|>\n你好<|end|>\n", text)
+        self.assertTrue(text.endswith("<|assistant|>\n"))
+        self.assertIn("message['role']", TEMPLATE)
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/test_r4_architectures.py

@@ -0,0 +1,65 @@
+import json
+import shutil
+import tempfile
+import unittest
+from pathlib import Path
+
+import apply_patches
+
+ROOT = Path(__file__).resolve().parents[1]
+FIXTURE = ROOT / "tests" / "fixtures" / "fake-model"
+WHITELIST = apply_patches.load_whitelist(ROOT / "patches" / "architectures" / "whitelist.txt")
+
+
+class ArchitectureTest(unittest.TestCase):
+    def setUp(self) -> None:
+        self.tmp = tempfile.TemporaryDirectory()
+        self.model = Path(self.tmp.name) / "model"
+        shutil.copytree(FIXTURE, self.model)
+
+    def tearDown(self) -> None:
+        self.tmp.cleanup()
+
+    def write_architectures(self, names: list[str]) -> None:
+        path = self.model / "config.json"
+        data = json.loads(path.read_text(encoding="utf-8"))
+        data["architectures"] = names
+        path.write_text(json.dumps(data), encoding="utf-8")
+
+    def test_whitelist_passes(self) -> None:
+        result = apply_patches.apply_architectures(self.model, WHITELIST, {}, "fail")
+        self.assertEqual(result["architectures"], ["LlamaForCausalLM"])
+        self.assertFalse(result["changed"])
+
+    def test_unknown_architecture_fails(self) -> None:
+        self.write_architectures(["ExampleForCausalLM"])
+        with self.assertRaises(apply_patches.PatchError) as ctx:
+            apply_patches.apply_architectures(self.model, WHITELIST, {}, "fail")
+        self.assertIn("ExampleForCausalLM", str(ctx.exception))
+
+    def test_unknown_architecture_skip_keeps_name(self) -> None:
+        self.write_architectures(["ExampleForCausalLM"])
+        result = apply_patches.apply_architectures(self.model, WHITELIST, {}, "skip")
+        self.assertEqual(result["skipped"], ["ExampleForCausalLM"])
+        data = json.loads((self.model / "config.json").read_text(encoding="utf-8"))
+        self.assertEqual(data["architectures"], ["ExampleForCausalLM"])
+
+    def test_alias_rewrites_to_whitelist(self) -> None:
+        self.write_architectures(["LlamaForCausalLMAlias"])
+        result = apply_patches.apply_architectures(
+            self.model,
+            WHITELIST,
+            {"LlamaForCausalLMAlias": "LlamaForCausalLM"},
+            "fail",
+        )
+        self.assertEqual(result["rewritten"], ["LlamaForCausalLMAlias->LlamaForCausalLM"])
+        data = json.loads((self.model / "config.json").read_text(encoding="utf-8"))
+        self.assertEqual(data["architectures"], ["LlamaForCausalLM"])
+
+    def test_shipped_aliases_are_empty(self) -> None:
+        aliases = apply_patches.load_aliases(ROOT / "patches" / "architectures" / "aliases.json")
+        self.assertEqual(aliases, {})
+
+
+if __name__ == "__main__":
+    unittest.main()