feat: MetaX deploy - MinerU /file_parse API, CJK fonts, SAM3, fronten…

.gitignore

@@ -253,6 +253,19 @@ outputs/
 
 fastapi_app/invite_codes.txt
 models/*
+!models/.gitkeep
+!models/README.md
+!models/RMBG-2.0/
+!models/sam3/
+!models/sam3-official/
+models/RMBG-2.0/*
+!models/RMBG-2.0/.gitkeep
+models/sam3/*
+!models/sam3/.gitkeep
+models/sam3-official/*
+!models/sam3-official/sam3/
+models/sam3-official/sam3/*
+!models/sam3-official/sam3/.gitkeep
 outputs/*
 tmps/*
 data/*

dataflow_agent/toolkits/multimodaltool/bg_tool.py

@@ -12,16 +12,16 @@
 # - get_svg_render_desc: 返回 SVG 渲染工具的说明文本
 # ================================================================
 # BRIA-RMBG 2.0 高质量抠图工具
-# - 模型：RMBG 2.0（ONNX）
-# - 依赖：onnxruntime, pillow, numpy
+# - 模型：RMBG 2.0（Transformers / ModelScope 目录）
+# - 依赖：transformers, pillow, numpy
 # ================================================================
 
 from __future__ import annotations
 
 import os
 import subprocess
+import sys
 from pathlib import Path
-import platform
 
 import numpy as np
 from PIL import Image, ImageFilter
@@ -31,57 +31,61 @@
 from dataflow_agent.logger import get_logger
 
 CURRENT_DIR = Path(__file__).resolve().parent
-# Allow override via env var (e.g. in Docker: RMBG_MODEL_PATH=/app/models/RMBG-2.0)
+PROJECT_ROOT = CURRENT_DIR.parents[2]
+# Allow override via env var (e.g. RMBG_MODEL_PATH=/app/models/RMBG-2.0)
 _env_model_path = os.environ.get("RMBG_MODEL_PATH")
-MODEL_PATH = Path(_env_model_path) if _env_model_path else CURRENT_DIR / "onnx" / "model.onnx"
+MODEL_PATH = Path(_env_model_path) if _env_model_path else PROJECT_ROOT / "models" / "RMBG-2.0"
 OUTPUT_DIR = CURRENT_DIR
 
 # 进程级抠图模型缓存：按 model_path 复用 BriaRMBG2Remover 实例
 _BG_RMBG_MODEL_CACHE: dict[str, "BriaRMBG2Remover"] = {}
 log = get_logger(__name__)
 
 
+def _has_rmbg_model(model_path: Path) -> bool:
+    return (model_path / "config.json").exists() and (model_path / "model.safetensors").exists()
+
+
 def ensure_model(model_path: Path) -> None:
     """
     确保本地存在 RMBG-2.0 模型权重。
 
-    若 ``model_path`` 对应的文件不存在，则通过 ModelScope 下载
-    ``AI-ModelScope/RMBG-2.0`` 模型到该路径所在目录。
+    当前实现期望 ``model_path`` 是一个 HuggingFace / ModelScope 风格的
+    模型目录；若目录缺失，则自动下载 ``AI-ModelScope/RMBG-2.0``。
 
     参数
     ----
     model_path:
-        本地模型文件路径（通常为 ONNX 或 transformers 权重）。
+        本地模型目录路径。
 
     异常
     ----
     FileNotFoundError
         当下载结束后仍未在 ``model_path`` 处找到模型文件时抛出。
     """
-    if model_path.exists():
+    if model_path.is_file():
+        model_path = model_path.parent
+
+    if _has_rmbg_model(model_path):
         log.info(f"模型已存在: {model_path}")
         return
 
-    log.info("未检测到模型文件，正在下载 RMBG-2.0 权重...")
-
-    # 确保目录存在
-    model_path.parent.mkdir(parents=True, exist_ok=True)
-    # 判断当前系统是否为Windows
-    is_windows = platform.system().lower() == "windows"
-    # Windows用双引号包裹路径，Linux/macOS用单引号（保持原有逻辑）
-    quote = '"' if is_windows else "'"
-    # 直接下载到目标目录
-    cmd = (
-        f"modelscope download "
-        f"--model AI-ModelScope/RMBG-2.0 "
-        f"--local_dir {quote}{model_path}{quote} "
-    )
-    os.system(cmd)
+    log.info("未检测到 RMBG 模型目录，正在下载 RMBG-2.0 权重...")
+
+    model_path.mkdir(parents=True, exist_ok=True)
+    download_code = (
+        "from modelscope import snapshot_download\n"
+        "snapshot_download("
+        "'AI-ModelScope/RMBG-2.0', "
+        "local_dir=r'''%s''', "
+        "allow_patterns=['*.json', '*.py', '*.safetensors']"
+        ")\n"
+    ) % str(model_path)
+    subprocess.run([sys.executable, "-c", download_code], check=True)
 
-    # 检查下载是否成功
-    if not model_path.exists():
+    if not _has_rmbg_model(model_path):
         raise FileNotFoundError(
-            f"模型下载失败：未找到 {model_path}。\n"
+            f"模型下载失败：目录 {model_path} 缺少 config.json 或 model.safetensors。\n"
             "请检查 ModelScope 或手动下载。"
         )
 
@@ -109,6 +113,8 @@ def __init__(self, model_path: Path | None = None, output_dir: Path | None = Non
             输出目录。若为 None，则使用当前文件所在目录。
         """
         self.model_path = Path(model_path) if model_path else MODEL_PATH
+        if self.model_path.is_file():
+            self.model_path = self.model_path.parent
         self.output_dir = Path(output_dir) if output_dir else OUTPUT_DIR
 
         ensure_model(self.model_path)

dataflow_agent/workflow/__init__.py

@@ -6,11 +6,37 @@
 from .paper2video_subprocess import run_paper2video_via_subprocess
 from .registry import RuntimeRegistry
 
-# ---- 1. 自动发现并导入所有工作流定义模块 ---------------------------------
 _pkg_path = Path(__file__).resolve().parent
-for py in _pkg_path.glob("wf_*.py"):
-    mod_name = f"{__name__}.{py.stem}"
-    importlib.import_module(mod_name)
+_imported_workflow_modules: set[str] = set()
+
+
+def _workflow_module_names() -> list[str]:
+    return sorted(
+        f"{__name__}.{py.stem}"
+        for py in _pkg_path.glob("wf_*.py")
+    )
+
+
+def _import_workflow_modules_until(name: str | None = None) -> None:
+    """
+    Lazy-load workflow definition modules.
+
+    Importing every workflow during FastAPI startup pulls in heavyweight
+    dependencies such as torchvision/transformers and makes the backend look
+    dead for a long time. Only import enough modules to satisfy the requested
+    workflow registration, and fall back to importing everything only when the
+    caller explicitly asks for the full list.
+    """
+    if name is not None and name in RuntimeRegistry._workflows:
+        return
+
+    for mod_name in _workflow_module_names():
+        if mod_name in _imported_workflow_modules:
+            continue
+        importlib.import_module(mod_name)
+        _imported_workflow_modules.add(mod_name)
+        if name is not None and name in RuntimeRegistry._workflows:
+            return
 
 # ---- 2. 工作流的统一接口 ---------------------------------------------
 def get_workflow(name: str):
@@ -23,6 +49,7 @@ def get_workflow(name: str):
     Returns:
         Callable: 用于构建该工作流图的工厂函数
     """
+    _import_workflow_modules_until(name)
     return RuntimeRegistry.get(name)
 
 
@@ -37,8 +64,9 @@ async def run_workflow(name: str, state):
 
 
 # ---- 3. 工作流注册信息公开接口 -------------------------------------------
-list_workflows = RuntimeRegistry.all
+def list_workflows():
+    _import_workflow_modules_until()
+    return RuntimeRegistry.all()
 
 # ---- 3. 工作流注册信息公开接口 -------------------------------------------
 # 提供所有已注册工作流的列表，便于外部查询与 introspection
-list_workflows = RuntimeRegistry.all

deploy/app_config.sh

@@ -3,8 +3,24 @@
 # Shared FastAPI runtime config for deploy scripts.
 # Environment variables can override these defaults.
 
+DEPLOY_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+PROJECT_ROOT="$(cd "$DEPLOY_DIR/.." && pwd)"
+
+APP_HOST="${APP_HOST:-0.0.0.0}"
 APP_PORT="${APP_PORT:-8000}"
-APP_WORKERS="${APP_WORKERS:-2}"
+APP_WORKERS="${APP_WORKERS:-1}"
 APP_CONDA_ENV="${APP_CONDA_ENV:-}"
-APP_PYTHON="${APP_PYTHON:-}"
-CONDA_SH="${CONDA_SH:-/root/miniconda3/etc/profile.d/conda.sh}"
+APP_PYTHON="${APP_PYTHON:-/opt/conda/bin/python}"
+APP_FALLBACK_PYTHON="${APP_FALLBACK_PYTHON:-/opt/conda/bin/python}"
+CONDA_SH="${CONDA_SH:-/opt/conda/etc/profile.d/conda.sh}"
+
+# Keep the legacy external repo as a fallback only.
+PAPER2ANY_ASSET_ROOT="${PAPER2ANY_ASSET_ROOT:-/mnt/paper2any/lz/github-proj/Paper2Any}"
+MODEL_SERVER_ENV_FILE="${MODEL_SERVER_ENV_FILE:-logs/model_servers.env}"
+
+SAM3_SERVER_URLS="${SAM3_SERVER_URLS:-http://127.0.0.1:8021}"
+# Leave model paths empty by default so deploy/start.sh can prefer repo-local models first.
+SAM3_HOME="${SAM3_HOME:-}"
+SAM3_CHECKPOINT_PATH="${SAM3_CHECKPOINT_PATH:-}"
+SAM3_BPE_PATH="${SAM3_BPE_PATH:-}"
+RMBG_MODEL_PATH="${RMBG_MODEL_PATH:-}"