feat: 增加失败任务重试与 CUDA 动态库路径注入

app/node_manager.py

@@ -15,7 +15,13 @@ from pathlib import Path
 
 from wov_sdk.models import InvokeRequest, InvokeResponse, NodeManifest
 
-from app.config import NODE_READY_TIMEOUT_SECONDS, NODE_REAP_INTERVAL_SECONDS, SDK_SRC, WORKSPACE_ROOT
+from app.config import (
+    NODE_INVOKE_TIMEOUT_SECONDS,
+    NODE_READY_TIMEOUT_SECONDS,
+    NODE_REAP_INTERVAL_SECONDS,
+    SDK_SRC,
+    WORKSPACE_ROOT,
+)
 from app.db import Database
 
 
@@ -33,6 +39,36 @@ def _idle_seconds(last_used_at: str) -> float:
         return 0.0
 
 
+def _cuda_library_dirs(repo_dir: Path) -> list[Path]:
+    venv = repo_dir / ".venv"
+    site_packages: list[Path] = []
+    windows_site = venv / "Lib" / "site-packages"
+    if windows_site.is_dir():
+        site_packages.append(windows_site)
+    site_packages.extend((venv / "lib").glob("python3*/site-packages"))
+
+    dirs: list[Path] = []
+    for site in site_packages:
+        for vendor in ("cublas", "cudnn"):
+            for subdir in ("lib", "bin"):
+                lib_dir = site / "nvidia" / vendor / subdir
+                if lib_dir.is_dir():
+                    dirs.append(lib_dir)
+    return dirs
+
+
+def _with_cuda_library_path(env: dict[str, str], repo_dir: Path) -> dict[str, str]:
+    dirs = [str(path) for path in _cuda_library_dirs(repo_dir)]
+    if not dirs:
+        return env
+    var = "PATH" if os.name == "nt" else "LD_LIBRARY_PATH"
+    existing = [item for item in env.get(var, "").split(os.pathsep) if item]
+    additions = [path for path in dirs if path not in existing]
+    if additions:
+        env[var] = os.pathsep.join(additions + existing)
+    return env
+
+
 @dataclass
 class NodeRuntime:
     instance_id: str
@@ -107,7 +143,8 @@ class NodeManager:
             item for item in [str(SDK_SRC), current_pythonpath] if item
         )
         env.update(manifest.env)
-        return env
+        repo_dir = WORKSPACE_ROOT / manifest.repo_dir
+        return _with_cuda_library_path(env, repo_dir)
 
     def _start_locked(self, manifest: NodeManifest) -> NodeRuntime:
         runtime = NodeRuntime(
@@ -270,7 +307,7 @@ class NodeManager:
             method="POST",
         )
         try:
-            with urllib.request.urlopen(http_request, timeout=300) as response:
+            with urllib.request.urlopen(http_request, timeout=NODE_INVOKE_TIMEOUT_SECONDS) as response:
                 payload = json.loads(response.read().decode("utf-8"))
             return InvokeResponse.from_dict(payload)
         except urllib.error.HTTPError as exc: