ci: split optional dependency profiles

.cursor/plans/20260810-11_拆分CI依赖.plan.md

@@ -0,0 +1,36 @@
+# 拆分 CI 依赖
+
+## 背景 / 目标
+
+当前 Python Tests 对所有改动都安装完整 Torch / Hugging Face / MCP 依赖，而完整测试执行仅约 14 秒。目标是在保持 `requirements.txt` 全功能安装合同和模型相关变更覆盖的前提下，缩短普通 PR 的依赖安装时间并消除重复触发。
+
+## 任务分解（checklist）
+
+- [ ] 拆分基础运行、Kronos 模型、MCP 与 CI 测试依赖。
+- [ ] 保留 `requirements.txt` 作为完整功能安装入口。
+- [ ] 普通 Python Tests 仅安装 CI 依赖，并启用 pip 缓存。
+- [ ] 限制常规触发为 main push 与面向 main 的 PR，避免 feature push + PR 重复运行。
+- [ ] 新增模型路径专用 workflow，安装模型依赖并验证导入。
+- [ ] 删除当前仓库无引用的 scikit-learn / lightgbm 依赖。
+- [ ] 更新 README 安装档位说明。
+- [ ] 在隔离环境验证基础安装不包含模型/MCP 包且 API 可启动、完整测试通过。
+- [ ] 提交、推送、创建 PR，等待 CI 通过后合并并关闭 Issue #5。
+
+## 验收标准
+
+- `requirements.txt` 递归安装基础、Kronos 与 MCP 全部功能依赖。
+- `requirements-ci.txt` 不包含 torch、Hugging Face、MCP、scikit-learn 或 lightgbm。
+- 普通 CI 在缺少可选包的环境中完成 API 启动与完整 pytest。
+- 模型代码或依赖变化会触发独立 model-import 检查。
+- workflow YAML 可解析，GitHub Actions 绿灯。
+
+## 风险与回滚
+
+- 风险：隐藏的顶层导入导致基础 CI 无法启动；用全新虚拟环境实际安装并启动 API 验证。
+- 风险：模型代码失去覆盖；以 paths 触发的专用 workflow 保留依赖和导入检查。
+- 回滚：恢复单一 requirements 与原 workflow 安装命令，不涉及数据库和生产数据。
+
+## 关键决策记录
+
+- 不把模型依赖从完整安装入口移除，只让普通 CI 使用轻量档位。
+- benchmark 不纳入普通 pytest；模型 workflow 只做确定性 import/构造检查，不下载远端权重。

.github/workflows/model-tests.yml

@@ -0,0 +1,67 @@
+name: Kronos Model Imports
+
+on:
+  push:
+    branches: [main]
+    paths:
+      - 'app/services/kronos_model/**'
+      - 'app/services/kronos_predict_service.py'
+      - 'app/services/first_limit_alpha_service.py'
+      - 'strategy/first_limit_alpha/train_sequence.py'
+      - 'tests/test_first_limit_alpha.py'
+      - 'requirements*.txt'
+      - '.github/workflows/model-tests.yml'
+  pull_request:
+    branches: [main]
+    paths:
+      - 'app/services/kronos_model/**'
+      - 'app/services/kronos_predict_service.py'
+      - 'app/services/first_limit_alpha_service.py'
+      - 'strategy/first_limit_alpha/train_sequence.py'
+      - 'tests/test_first_limit_alpha.py'
+      - 'requirements*.txt'
+      - '.github/workflows/model-tests.yml'
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+concurrency:
+  group: kronos-imports-${{ github.event.pull_request.number || github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  import-model-runtime:
+    runs-on: ubuntu-24.04
+
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Setup Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: '3.11'
+          cache: pip
+          cache-dependency-path: |
+            requirements*.txt
+
+      - name: Install model runtime
+        run: |
+          python -m pip install --upgrade pip
+          pip install -r requirements-ci.txt -r requirements-kronos.txt
+          pip check
+
+      - name: Import Kronos runtime
+        run: |
+          python - <<'PY'
+          from app.services.kronos_model import Kronos, KronosPredictor, KronosTokenizer
+
+          assert Kronos is not None
+          assert KronosPredictor is not None
+          assert KronosTokenizer is not None
+          print("Kronos runtime imports: ok")
+          PY
+
+      - name: Test Torch sequence training
+        run: python -m pytest tests/test_first_limit_alpha.py -q -k sequence

.github/workflows/tests.yml

@@ -2,13 +2,21 @@ name: Python Tests
 
 on:
   push:
-    branches: ["**"]
+    branches: [main]
   pull_request:
-    branches: ["**"]
+    branches: [main]
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+concurrency:
+  group: python-tests-${{ github.event.pull_request.number || github.ref }}
+  cancel-in-progress: true
 
 jobs:
   test:
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-24.04
 
     steps:
       - name: Checkout
@@ -18,12 +26,15 @@ jobs:
         uses: actions/setup-python@v5
         with:
           python-version: '3.11'
+          cache: pip
+          cache-dependency-path: |
+            requirements*.txt
 
       - name: Install dependencies
         run: |
           python -m pip install --upgrade pip
-          pip install -r requirements.txt
-          pip install pytest
+          pip install -r requirements-ci.txt
+          pip check
 
       - name: Start API for regression tests
         env:

README.md

@@ -513,9 +513,22 @@ Alpha/
 ### 安装依赖
 
 ```bash
+# 完整安装：基础 API / 数据源 + Kronos + MCP + 训练后端
 pip3 install -r requirements.txt
 ```
 
+也可以按部署角色选择更小的依赖档位：
+
+| 安装档位 | 命令 | 用途 |
+|----------|------|------|
+| 基础运行 | `pip3 install -r requirements-base.txt` | FastAPI、行情数据、策略与本地缓存；Kronos / MCP 保持惰性可选 |
+| Kronos | `pip3 install -r requirements-base.txt -r requirements-kronos.txt` | 增加 Torch、Hugging Face 与 Kronos 推理能力 |
+| MCP | `pip3 install -r requirements-base.txt -r requirements-mcp.txt` | 增加独立 MCP Server |
+| 策略训练 | `pip3 install -r requirements-base.txt -r requirements-training.txt` | 增加 FirstLimit Alpha 首选的 LightGBM 训练后端；未安装时自动使用 RandomForest |
+| 完整功能 | `pip3 install -r requirements.txt` | 安装上述全部能力 |
+
+CI 使用 `requirements-ci.txt` 执行基础单元/API 回归；只有模型代码或模型依赖变化时，才运行独立的 Kronos 导入与 Torch 序列训练测试。模型权重仍在第一次真实预测时按需下载，不进入仓库或 CI 缓存。
+
 ### 启动服务
 
 ```bash
@@ -553,7 +566,6 @@ pip3 install -r requirements.txt
 
 日志文件：默认 `logs/server-18890.log`，随 `PORT` 变化。
 
-
 ### 行情缓存维护
 
 历史解析缺陷或错误补数可能污染 `data/market_kline.db`。维护命令默认只做只读预检；只有显式传入 `--apply` 才会先通过 SQLite backup API 创建并验证完整备份，再用单一事务清理早于 2000 年或无法解析的交易日期及其关联同步状态。

app/services/first_limit_alpha_service.py

@@ -25,7 +25,6 @@
     TrainingConfig,
 )
 from strategy.first_limit_alpha.feature_store import build_version_dir, latest_child, read_dataframe, read_json, write_json
-from strategy.first_limit_alpha.train_sequence import FirstLimitSequenceTrainer
 
 DEFAULT_GRAPHIC_CONFIG: dict[str, Any] = {
     "threshold_candidate": 6.0,
@@ -168,6 +167,10 @@ def train_baseline(self, training_cfg: TrainingConfig | None = None) -> dict[str
         return result
 
     def train_sequence(self, sequence_cfg: SequenceConfig | None = None) -> dict[str, Any]:
+        # Keep the core API usable without the optional Torch profile. Sequence
+        # training imports its backend only when this endpoint is invoked.
+        from strategy.first_limit_alpha.train_sequence import FirstLimitSequenceTrainer
+
         samples = self._load_latest_samples()
         out_dir = build_version_dir(self.root, "sequence_models", prefix="gru")
         baseline_meta: dict[str, Any] = {}

requirements-base.txt

@@ -0,0 +1,13 @@
+# Core API and market-data runtime.
+akshare>=1.16.0
+tushare>=1.4.0
+fastapi>=0.115.0
+uvicorn[standard]>=0.30.0
+pandas>=2.2.0
+pydantic>=2.8.0
+requests>=2.31.0
+httpx>=0.27.0
+tqdm>=4.60.0
+numpy>=1.24.0
+joblib>=1.4.0
+scikit-learn>=1.5.0

requirements-ci.txt

@@ -0,0 +1,3 @@
+# Deterministic unit/API regression environment; optional model/MCP imports stay lazy.
+-r requirements-base.txt
+pytest>=8.0.0

requirements-kronos.txt

@@ -0,0 +1,5 @@
+# Optional Kronos model runtime. Model weights are downloaded lazily.
+torch>=2.0.0
+einops>=0.8.1
+huggingface_hub>=0.33.0
+safetensors>=0.6.0

requirements-mcp.txt

@@ -0,0 +1,2 @@
+# Optional standalone MCP server runtime.
+mcp>=1.27.0

requirements-training.txt

@@ -0,0 +1,2 @@
+# Optional preferred backend for FirstLimit Alpha training.
+lightgbm>=4.3.0

requirements.txt

@@ -1,17 +1,5 @@
-akshare>=1.16.0
-tushare>=1.4.0
-fastapi>=0.115.0
-uvicorn[standard]>=0.30.0
-pandas>=2.2.0
-pydantic>=2.8.0
-requests>=2.31.0
-httpx>=0.27.0
-torch>=2.0.0
-einops>=0.8.1
-huggingface_hub>=0.33.0
-safetensors>=0.6.0
-tqdm>=4.60.0
-numpy>=1.24.0
-mcp>=1.27.0
-scikit-learn>=1.5.0
-lightgbm>=4.3.0
+# Full Alpha installation: API/data runtime + Kronos + MCP + training backends.
+-r requirements-base.txt
+-r requirements-kronos.txt
+-r requirements-mcp.txt
+-r requirements-training.txt

tests/test_dependency_profiles.py

@@ -0,0 +1,63 @@
+from pathlib import Path
+
+
+ROOT = Path(__file__).resolve().parents[1]
+
+
+def _requirement_lines(name: str) -> set[str]:
+    return {
+        line.strip()
+        for line in (ROOT / name).read_text(encoding="utf-8").splitlines()
+        if line.strip() and not line.lstrip().startswith("#")
+    }
+
+
+def test_full_install_includes_all_supported_profiles():
+    assert _requirement_lines("requirements.txt") == {
+        "-r requirements-base.txt",
+        "-r requirements-kronos.txt",
+        "-r requirements-mcp.txt",
+        "-r requirements-training.txt",
+    }
+
+
+def test_ci_profile_excludes_optional_heavy_dependencies():
+    ci_lines = _requirement_lines("requirements-ci.txt")
+    base_lines = _requirement_lines("requirements-base.txt")
+    combined = "\n".join(sorted(ci_lines | base_lines)).lower()
+
+    assert "-r requirements-base.txt" in ci_lines
+    assert "pytest>=8.0.0" in ci_lines
+    for package in (
+        "torch",
+        "einops",
+        "huggingface_hub",
+        "safetensors",
+        "mcp",
+        "lightgbm",
+    ):
+        assert package not in combined
+
+
+def test_model_workflow_tracks_model_paths_and_dependencies():
+    workflow = (ROOT / ".github/workflows/model-tests.yml").read_text(
+        encoding="utf-8"
+    )
+
+    assert "app/services/kronos_model/**" in workflow
+    assert "app/services/kronos_predict_service.py" in workflow
+    assert "strategy/first_limit_alpha/train_sequence.py" in workflow
+    assert "requirements*.txt" in workflow
+    assert "requirements-ci.txt" in workflow
+    assert "requirements-kronos.txt" in workflow
+    assert "from app.services.kronos_model import" in workflow
+    assert "-k sequence" in workflow
+
+
+def test_core_service_lazy_loads_torch_sequence_backend():
+    service = (ROOT / "app/services/first_limit_alpha_service.py").read_text(
+        encoding="utf-8"
+    )
+
+    import_line = "from strategy.first_limit_alpha.train_sequence import FirstLimitSequenceTrainer"
+    assert service.index(import_line) > service.index("def train_sequence(")

tests/test_first_limit_alpha.py

@@ -3,14 +3,14 @@
 from pathlib import Path
 
 import pandas as pd
+import pytest
 
 from app.services.first_limit_alpha_service import FirstLimitAlphaService
 from app.services.kline_store import KlineSQLiteStore
 from strategy.first_limit_alpha.data_builder import FirstLimitAlphaDataBuilder
 from strategy.first_limit_alpha.features import FirstLimitFeatureBuilder
 from strategy.first_limit_alpha.modeling import FirstLimitBaselineTrainer
 from strategy.first_limit_alpha.schema import LabelConfig, SampleBuildConfig, SequenceConfig, TrainingConfig
-from strategy.first_limit_alpha.train_sequence import FirstLimitSequenceTrainer
 
 
 def _symbol_rows(base_close: float, event_mode: str, start: str = "2026-01-01") -> list[dict]:
@@ -181,6 +181,9 @@ def test_baseline_trainer_excludes_future_columns(tmp_path: Path):
 
 
 def test_sequence_trainer_small_sample_returns_graceful_result(tmp_path: Path):
+    pytest.importorskip("torch")
+    from strategy.first_limit_alpha.train_sequence import FirstLimitSequenceTrainer
+
     store = KlineSQLiteStore(str(tmp_path / "market_kline.db"))
     store.upsert_symbol_klines("600001", _symbol_rows(10.0, "strong"), "2026-04-22T11:00:00+08:00")
     samples = pd.DataFrame(
@@ -200,6 +203,9 @@ def test_sequence_trainer_small_sample_returns_graceful_result(tmp_path: Path):
 
 
 def test_sequence_trainer_outputs_metrics_and_predictions(tmp_path: Path):
+    pytest.importorskip("torch")
+    from strategy.first_limit_alpha.train_sequence import FirstLimitSequenceTrainer
+
     store = KlineSQLiteStore(str(tmp_path / "market_kline.db"))
     name_map = {}
     for idx in range(40):