feat: 增加失败任务重试与 CUDA 动态库路径注入

tests/test_node_manager.py

@@ -1,3 +1,4 @@
+import os
 import subprocess
 import sys
 import time
@@ -7,7 +8,14 @@ from pathlib import Path
 import pytest
 
 from app.db import Database
-from app.node_manager import NodeManager, NodeRuntime, _idle_seconds, _now_iso
+from app.node_manager import (
+    NodeManager,
+    NodeRuntime,
+    _cuda_library_dirs,
+    _idle_seconds,
+    _now_iso,
+    _with_cuda_library_path,
+)
 from wov_sdk.models import InvokeRequest, NodeManifest
 
 WORKSPACE = Path(__file__).resolve().parent.parent.parent
@@ -86,6 +94,130 @@ def test_resolve_command_prefers_node_venv(db: Database, tmp_path, monkeypatch)
     assert manager._resolve_command(manifest)[0] == str(unix_python)
 
 
+def test_cuda_library_dirs_finds_unix_nvidia_libs(tmp_path) -> None:
+    cublas = (
+        tmp_path / ".venv" / "lib" / "python3.12" / "site-packages" / "nvidia" / "cublas" / "lib"
+    )
+    cudnn = (
+        tmp_path / ".venv" / "lib" / "python3.12" / "site-packages" / "nvidia" / "cudnn" / "lib"
+    )
+    cublas.mkdir(parents=True)
+    cudnn.mkdir(parents=True)
+
+    dirs = _cuda_library_dirs(tmp_path)
+
+    assert cublas in dirs
+    assert cudnn in dirs
+
+
+def test_cuda_library_dirs_finds_windows_nvidia_bins(tmp_path) -> None:
+    cublas = tmp_path / ".venv" / "Lib" / "site-packages" / "nvidia" / "cublas" / "bin"
+    cublas.mkdir(parents=True)
+
+    dirs = _cuda_library_dirs(tmp_path)
+
+    assert cublas in dirs
+
+
+def test_with_cuda_library_path_prepends_on_posix(monkeypatch) -> None:
+    monkeypatch.setattr(os, "name", "posix")
+    monkeypatch.setattr(
+        "app.node_manager._cuda_library_dirs",
+        lambda repo_dir: ["/opt/nvidia/cublas/lib", "/opt/nvidia/cudnn/lib"],
+    )
+
+    env = _with_cuda_library_path({"LD_LIBRARY_PATH": "/usr/lib/foo"}, Path("/repo"))
+
+    entries = env["LD_LIBRARY_PATH"].split(os.pathsep)
+    assert entries[:2] == ["/opt/nvidia/cublas/lib", "/opt/nvidia/cudnn/lib"]
+    assert entries[-1] == "/usr/lib/foo"
+
+
+def test_with_cuda_library_path_uses_path_on_windows(monkeypatch) -> None:
+    monkeypatch.setattr(os, "name", "nt")
+    monkeypatch.setattr(
+        "app.node_manager._cuda_library_dirs",
+        lambda repo_dir: ["/opt/nvidia/cublas/bin"],
+    )
+
+    env = _with_cuda_library_path({"PATH": "/usr/local/bin"}, Path("/repo"))
+
+    entries = env["PATH"].split(os.pathsep)
+    assert entries[0] == "/opt/nvidia/cublas/bin"
+    assert entries[-1] == "/usr/local/bin"
+
+
+def test_with_cuda_library_path_without_existing(monkeypatch) -> None:
+    monkeypatch.setattr(os, "name", "posix")
+    monkeypatch.setattr(
+        "app.node_manager._cuda_library_dirs",
+        lambda repo_dir: ["/opt/nvidia/cublas/lib"],
+    )
+
+    env = _with_cuda_library_path({}, Path("/repo"))
+
+    assert env["LD_LIBRARY_PATH"] == "/opt/nvidia/cublas/lib"
+
+
+def test_with_cuda_library_path_deduplicates(monkeypatch) -> None:
+    monkeypatch.setattr(os, "name", "posix")
+    monkeypatch.setattr(
+        "app.node_manager._cuda_library_dirs",
+        lambda repo_dir: ["/opt/nvidia/cublas/lib", "/opt/nvidia/cudnn/lib"],
+    )
+
+    env = _with_cuda_library_path(
+        {"LD_LIBRARY_PATH": "/opt/nvidia/cublas/lib:/usr/lib/foo"},
+        Path("/repo"),
+    )
+
+    entries = env["LD_LIBRARY_PATH"].split(os.pathsep)
+    assert entries == ["/opt/nvidia/cudnn/lib", "/opt/nvidia/cublas/lib", "/usr/lib/foo"]
+
+
+def test_with_cuda_library_path_all_present(monkeypatch) -> None:
+    monkeypatch.setattr(os, "name", "posix")
+    monkeypatch.setattr(
+        "app.node_manager._cuda_library_dirs",
+        lambda repo_dir: ["/opt/nvidia/cublas/lib"],
+    )
+    original = {"LD_LIBRARY_PATH": "/opt/nvidia/cublas/lib:/usr/lib/foo"}
+
+    env = _with_cuda_library_path(dict(original), Path("/repo"))
+
+    assert env["LD_LIBRARY_PATH"] == original["LD_LIBRARY_PATH"]
+
+
+def test_with_cuda_library_path_without_libs(monkeypatch) -> None:
+    monkeypatch.setattr("app.node_manager._cuda_library_dirs", lambda repo_dir: [])
+
+    env = _with_cuda_library_path({"LD_LIBRARY_PATH": "/usr/lib/foo"}, Path("/repo"))
+
+    assert env == {"LD_LIBRARY_PATH": "/usr/lib/foo"}
+
+
+def test_node_env_injects_cuda_library_path(db: Database, tmp_path, monkeypatch) -> None:
+    monkeypatch.setattr("app.node_manager.WORKSPACE_ROOT", tmp_path)
+    repo = tmp_path / "wov-node-whisper"
+    cublas = repo / ".venv" / "lib" / "python3.12" / "site-packages" / "nvidia" / "cublas" / "lib"
+    cublas.mkdir(parents=True)
+    manifest = NodeManifest(
+        id="x",
+        name="x",
+        version="1",
+        capability="x",
+        command=["python", "-m", "x"],
+        repo_dir="wov-node-whisper",
+        env={"WOV_NODE_PORT": "0"},
+    )
+
+    env = NodeManager(db)._node_env(manifest)
+
+    assert str(cublas) in env["LD_LIBRARY_PATH"].split(os.pathsep)
+    assert env["WOV_NODE_PORT"] == "0"
+    assert "wov-sdk" in env["PYTHONPATH"]
+
+
 def test_acquire_unregistered(db: Database) -> None:
     manager = NodeManager(db)
     with pytest.raises(ValueError):
@@ -266,6 +398,51 @@ def test_invoke_network_error(db: Database, tmp_path) -> None:
         manager.shutdown()
 
 
+def test_invoke_uses_configured_timeout(db: Database, monkeypatch) -> None:
+    _register(db, echo_manifest())
+    manager = NodeManager(db)
+    captured = {}
+
+    class FakeRuntime:
+        instance_id = "ni_x"
+        status = "ready"
+        busy_count = 0
+
+    class FakeResponse:
+        def read(self) -> bytes:
+            return b'{"status":"completed","outputs":{"text":"ok"}}'
+
+        def __enter__(self):
+            return self
+
+        def __exit__(self, *args) -> bool:
+            return False
+
+    def fake_acquire(node_id):
+        return FakeRuntime(), "http://127.0.0.1:1"
+
+    def fake_urlopen(request, timeout):
+        captured["timeout"] = timeout
+        return FakeResponse()
+
+    monkeypatch.setattr("app.node_manager.NODE_INVOKE_TIMEOUT_SECONDS", 123)
+    monkeypatch.setattr(manager, "acquire", fake_acquire)
+    monkeypatch.setattr("app.node_manager.urllib.request.urlopen", fake_urlopen)
+
+    response = manager.invoke(
+        "echo",
+        InvokeRequest(
+            run_id="run_t",
+            node_instance_id="",
+            inputs={},
+            output_dir=".",
+        ),
+    )
+
+    assert response.status == "completed"
+    assert captured["timeout"] == 123
+
+
 def test_release_unknown(db: Database) -> None:
     manager = NodeManager(db)
     manager.release("missing")