Skip to content

利用 AI 大模型和自动化工作流,根据主题或关键词一键生成高清短视频。Generate HD short videos from a topic or keyword with an automated AI workflow.

fix(llm): close SDK clients and streams on every outcome

Rudy Celeklia8393b717002
diff --git a/app/services/llm.py b/app/services/llm.py
index 5bab0fc7..7c310af0 100644
--- a/app/services/llm.py
+++ b/app/services/llm.py
@@ -255,6 +255,8 @@ def _extract_qwen_generation_text(response) -> str:
def _generate_response(prompt: str, app_config=None) -> str:
+ sdk_client = None
+ sdk_stream = None
try:
# WebUI 在视频生成期间允许用户准备下一条文案。调用方可以传入提交瞬间
# 的配置快照,确保模型请求重试期间不会因为后台任务结束并应用新配置,
@@ -410,7 +412,7 @@ def _generate_response(prompt: str, app_config=None) -> str:
# Cloudflare 当前推荐的 AI Gateway REST API 兼容 OpenAI SDK。
# Account ID 用于构造统一端点,Gateway ID 通过请求头选择;这里
# 不再调用 Workers AI 的 /ai/run/{model} 专用接口。
- client = OpenAI(
+ client = sdk_client = OpenAI(
api_key=api_key,
base_url=(
f"https://api.cloudflare.com/client/v4/accounts/{account_id}/ai/v1"
@@ -450,7 +452,7 @@ def _generate_response(prompt: str, app_config=None) -> str:
# 这里在 Azure 分支内完成请求并立即返回,避免客户端被后续 fallback
# 覆盖,导致用户配置的 Azure 凭证通过校验但实际请求没有被使用。
logger.info(f"requesting azure chat completion, model: {model_name}")
- client = AzureOpenAI(
+ client = sdk_client = AzureOpenAI(
api_key=api_key,
api_version=api_version,
azure_endpoint=base_url,
@@ -594,11 +596,11 @@ def _generate_response(prompt: str, app_config=None) -> str:
if adapter == "modelscope":
content = ""
- client = OpenAI(
+ client = sdk_client = OpenAI(
api_key=api_key,
base_url=base_url,
)
- response = client.chat.completions.create(
+ response = sdk_stream = client.chat.completions.create(
model=model_name,
messages=[{"role": "user", "content": prompt}],
extra_body={"enable_thinking": False},
@@ -619,7 +621,7 @@ def _generate_response(prompt: str, app_config=None) -> str:
else:
raise Exception(f"[{llm_provider}] returned an empty response")
- client = OpenAI(
+ client = sdk_client = OpenAI(
api_key=api_key,
base_url=base_url,
)
@@ -643,6 +645,17 @@ def _generate_response(prompt: str, app_config=None) -> str:
except Exception as e:
return f"Error: {_sanitize_error_message(e)}"
+ finally:
+ for resource in (sdk_stream, sdk_client):
+ close = getattr(resource, "close", None)
+ if callable(close):
+ try:
+ close()
+ except Exception as cleanup_error:
+ logger.warning(
+ f"could not close LLM transport: {type(cleanup_error).__name__}"
+ )
+
def test_connection() -> tuple[bool, str, float]:
"""
diff --git a/test/services/test_llm_client_cleanup.py b/test/services/test_llm_client_cleanup.py
new file mode 100644
index 00000000..6874b23b
--- /dev/null
+++ b/test/services/test_llm_client_cleanup.py
@@ -0,0 +1,42 @@
+import unittest
+from unittest.mock import patch
+import httpx
+from openai import OpenAI, AzureOpenAI
+from app.services import llm
+
+
+class TestLlmClientCleanup(unittest.TestCase):
+ def test_sdk_transport_closed_after_success_and_request_failure(self):
+ cases = [
+ ("openai", "OpenAI", {"openai_api_key": "test-key", "openai_model_name": "test-model", "openai_base_url": "https://example.test/v1"}),
+ ("azure", "AzureOpenAI", {"azure_api_key": "test-key", "azure_model_name": "test-model", "azure_base_url": "https://example.test", "azure_api_version": "2024-02-15-preview"}),
+ ("cloudflare", "OpenAI", {"cloudflare_api_key": "test-key", "cloudflare_model_name": "test-model", "cloudflare_account_id": "test-account", "cloudflare_gateway_id": "test-gateway"}),
+ ]
+ for provider, constructor, settings in cases:
+ for failed in (False, True):
+ with self.subTest(provider=provider, failed=failed):
+ def request(req):
+ if failed:
+ raise httpx.ReadTimeout("lost response", request=req)
+ return httpx.Response(200, json={"id": "test", "object": "chat.completion", "created": 1, "model": "test-model", "choices": [{"index": 0, "finish_reason": "stop", "message": {"role": "assistant", "content": "valid narration"}}]})
+ transport = httpx.Client(transport=httpx.MockTransport(request))
+ cls = AzureOpenAI if constructor == "AzureOpenAI" else OpenAI
+ def create(**kwargs):
+ return cls(**kwargs, max_retries=0, http_client=transport)
+ with patch.object(llm, constructor, side_effect=create):
+ result = llm._generate_response("hello", app_config={"llm_provider": provider, **settings})
+ self.assertEqual(result.startswith("Error:"), failed, result)
+ self.assertTrue(transport.is_closed)
+
+ def test_stream_transport_closed_after_partial_stream_error(self):
+ class BrokenStream(httpx.SyncByteStream):
+ def __iter__(self):
+ yield b'data: {"id":"test","object":"chat.completion.chunk","created":1,"model":"test-model","choices":[{"index":0,"delta":{"content":"partial"}}]}\n\n'
+ raise httpx.ReadError("lost SSE connection")
+ transport = httpx.Client(transport=httpx.MockTransport(lambda req: httpx.Response(200, headers={"content-type": "text/event-stream"}, stream=BrokenStream())))
+ def create(**kwargs):
+ return OpenAI(**kwargs, max_retries=0, http_client=transport)
+ with patch.object(llm, "OpenAI", side_effect=create):
+ result = llm._generate_response("hello", app_config={"llm_provider": "modelscope", "modelscope_api_key": "test-key", "modelscope_model_name": "test-model", "modelscope_base_url": "https://example.test/v1"})
+ self.assertTrue(result.startswith("Error:"), result)
+ self.assertTrue(transport.is_closed)