chore: import upstream snapshot with attribution

tests/llm/nvidia_impl/test_nvidia_client_cleanup.py

@@ -0,0 +1,69 @@
+"""Regression tests for nvidia_openai_embed AsyncOpenAI cleanup.
+
+Mirrors the cleanup contract established for the Anthropic backend
+(``tests/llm/anthropic_impl/test_anthropic_client_cleanup.py``, PR #3261):
+the per-call ``AsyncOpenAI`` client must be released on both the success and
+error paths so its httpx connection pool does not leak one pool per call.
+
+``nvidia_openai_embed`` now holds the client in an ``async with`` (matching
+``openai_embed``); these tests pin that behavior. The raw function is invoked
+via ``.func.__wrapped__`` to bypass the ``EmbeddingFunc`` and tenacity
+``@retry`` wrappers.
+"""
+
+from types import SimpleNamespace
+from unittest.mock import AsyncMock, patch
+
+import pytest
+
+from lightrag.llm.nvidia_openai import nvidia_openai_embed
+
+
+class _FakeAsyncOpenAI:
+    """Mimics ``AsyncOpenAI`` as an async context manager.
+
+    The real SDK's ``__aexit__`` closes the client; this fake does the same so
+    the test can assert on ``close()``.
+    """
+
+    def __init__(self, *, create):
+        self.embeddings = SimpleNamespace(create=create)
+        self.close = AsyncMock()
+
+    async def __aenter__(self):
+        return self
+
+    async def __aexit__(self, exc_type, exc, tb):
+        await self.close()
+
+
+def _make_response(embedding=(0.1, 0.2, 0.3)):
+    return SimpleNamespace(data=[SimpleNamespace(embedding=list(embedding))])
+
+
+@pytest.mark.offline
+@pytest.mark.asyncio
+async def test_client_closed_after_success():
+    """Successful embed: client.close() runs after the result is returned."""
+    fake = _FakeAsyncOpenAI(create=AsyncMock(return_value=_make_response()))
+
+    with patch("lightrag.llm.nvidia_openai.AsyncOpenAI", return_value=fake):
+        result = await nvidia_openai_embed.func.__wrapped__(texts=["hello"])
+
+    assert result.tolist() == [[0.1, 0.2, 0.3]]
+    fake.embeddings.create.assert_awaited_once()
+    fake.close.assert_awaited()
+
+
+@pytest.mark.offline
+@pytest.mark.asyncio
+async def test_client_closed_on_error():
+    """API error: client.close() runs before the error propagates."""
+    fake = _FakeAsyncOpenAI(create=AsyncMock(side_effect=RuntimeError("boom")))
+
+    with patch("lightrag.llm.nvidia_openai.AsyncOpenAI", return_value=fake):
+        with pytest.raises(RuntimeError, match="boom"):
+            await nvidia_openai_embed.func.__wrapped__(texts=["hello"])
+
+    fake.embeddings.create.assert_awaited_once()
+    fake.close.assert_awaited()