初始提交：天垓150兼容引擎 I0–I4/I6/I7

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[@]}"