| |
| """镜像入口：镜像不带权重，启动前从 Hugging Face 拉取模型，然后 exec uvicorn。 |
| |
| - LAYA_MODEL_VERSION：HF tag/分支（默认 v2） |
| - LAYA_HF_ENDPOINT：模型源（默认 https://huggingface.co，可设 https://hf-mirror.com） |
| - LAYA_PRELOAD：显式预载精度，如 fp8,int4 |
| - 未设 LAYA_PRELOAD 且本地无模型时，自动下载 LAYA_MODEL_PRECISION（默认 fp8）； |
| LAYA_AUTO_DOWNLOAD=0 可关闭自动下载（改为管理员经 /internal/models/catalog 手动下载） |
| """ |
| import os |
| import time |
|
|
| from server.model_manager import ModelManager |
|
|
|
|
| def main(): |
| mm = ModelManager(os.environ.get("LAYA_MODEL_DIR", "/data/models/ad-clf"), |
| os.environ.get("LAYA_DEVICE") or None) |
| preload = [p.strip() for p in os.environ.get("LAYA_PRELOAD", "").split(",") if p.strip()] |
| if not preload and not mm.installed() and os.environ.get("LAYA_AUTO_DOWNLOAD", "1") != "0": |
| preload = [os.environ.get("LAYA_MODEL_PRECISION", "fp8").strip() or "fp8"] |
|
|
| for p in preload: |
| r = mm.download(p) |
| print(f"[entrypoint] preload {p}: {r}", flush=True) |
| if not r.get("accepted"): |
| continue |
| while mm.status()["state"] == "downloading": |
| time.sleep(1) |
| st = mm.status() |
| print(f"[entrypoint] preload {p} -> {st['state']}", flush=True) |
| if st.get("error"): |
| print(f"[entrypoint] error: {st['error']}", flush=True) |
|
|
| os.execvp("uvicorn", ["uvicorn", "server.main:app", "--host", "0.0.0.0", |
| "--port", "8080", "--workers", "1", "--proxy-headers"]) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
