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): fail fast with actionable hints when testing model connection

fengqiaoquf44414165fdb
diff --git a/app/services/llm.py b/app/services/llm.py
index 924b9753..b9008a1c 100644
--- a/app/services/llm.py
+++ b/app/services/llm.py
@@ -22,6 +22,24 @@ MIN_SCRIPT_PARAGRAPH_NUMBER = 1
MAX_SCRIPT_PARAGRAPH_NUMBER = 10
MAX_SCRIPT_PROMPT_LENGTH = 2000
MAX_SCRIPT_SYSTEM_PROMPT_LENGTH = 8000
+# “测试模型链接”只发送一条极小请求。OpenAI SDK 默认 600 秒超时并自动重试
+# 2 次,端点不可达(代理、防火墙或 DNS 故障)时,测试按钮会先长时间挂起,
+# 最后只留下一句 "Request timed out."(见 issue #1642)。这里给连接测试一个
+# 有界的短超时并禁用 SDK 重试,让失败快速、可诊断;正式的脚本生成链路保持
+# SDK 默认行为不变,避免误伤慢速推理模型和本地 Ollama 冷启动。
+LLM_CONNECTION_TEST_TIMEOUT_SECONDS = 30.0
+LLM_CONNECTION_TEST_MAX_RETRIES = 0
+# 这些片段来自 OpenAI SDK 的 APITimeoutError / APIConnectionError 以及
+# httpx 的底层连接异常文案;命中任意一个即按“端点不可达”处理。
+_CONNECTION_FAILURE_MARKERS = (
+ "timed out",
+ "timeout",
+ "connection error",
+ "connection refused",
+ "connection reset",
+ "connection aborted",
+ "failed to connect",
+)
_THINK_BLOCK_RE = re.compile(r"<think\b[^>]*>.*?</think>", re.IGNORECASE | re.DOTALL)
_UNCLOSED_THINK_BLOCK_RE = re.compile(r"<think\b[^>]*>.*$", re.IGNORECASE | re.DOTALL)
_URL_USERINFO_RE = re.compile(
@@ -228,6 +246,39 @@ def _get_response_field(value, key: str):
return getattr(value, key, None)
+def _sdk_transport_options(request_timeout, max_retries) -> dict:
+ """把可选的超时/重试约束转换为 OpenAI SDK 客户端构造参数。
+
+ 两个参数传 None 时返回空字典,客户端继续使用 SDK 默认值(600 秒超时、
+ 2 次自动重试),因此脚本生成等常规链路行为保持不变;只有连接测试这类
+ 需要快速失败的调用方显式传入有界值时,请求预算才会被收紧。
+ """
+ options = {}
+ if request_timeout is not None:
+ options["timeout"] = request_timeout
+ if max_retries is not None:
+ options["max_retries"] = max_retries
+ return options
+
+
+def _explain_connection_test_failure(error_message: str) -> str:
+ """为超时/连接类失败补充排查方向,而不是只显示一句 SDK 原始文案。
+
+ issue #1642 中用户的 API 本身可用,失败原因是本机到端点的链路问题
+ (代理、防火墙或 DNS),但 "Request timed out." 没有指向任何可能原因。
+ 其它错误(鉴权、配额等)保持原文返回。
+ """
+ normalized = (error_message or "").lower()
+ if not any(marker in normalized for marker in _CONNECTION_FAILURE_MARKERS):
+ return error_message
+ return (
+ f"{error_message} (the request did not complete within "
+ f"{LLM_CONNECTION_TEST_TIMEOUT_SECONDS:.0f}s). Please check that the "
+ "configured base URL is reachable from this machine, and that no proxy "
+ "or firewall is blocking the request."
+ )
+
+
def _extract_qwen_generation_text(response) -> str:
"""
从 DashScope Generation 响应中提取文本。
@@ -254,7 +305,12 @@ def _extract_qwen_generation_text(response) -> str:
return _normalize_text_response(text, "qwen")
-def _generate_response(prompt: str, app_config=None) -> str:
+def _generate_response(
+ prompt: str,
+ app_config=None,
+ request_timeout: float | None = None,
+ max_retries: int | None = None,
+) -> str:
sdk_client = None
sdk_stream = None
try:
@@ -418,6 +474,7 @@ def _generate_response(prompt: str, app_config=None) -> str:
f"https://api.cloudflare.com/client/v4/accounts/{account_id}/ai/v1"
),
default_headers={"cf-aig-gateway-id": gateway_id},
+ **_sdk_transport_options(request_timeout, max_retries),
)
response = client.chat.completions.create(
model=model_name,
@@ -456,6 +513,7 @@ def _generate_response(prompt: str, app_config=None) -> str:
api_key=api_key,
api_version=api_version,
azure_endpoint=base_url,
+ **_sdk_transport_options(request_timeout, max_retries),
)
response = client.chat.completions.create(
model=model_name, messages=[{"role": "user", "content": prompt}]
@@ -631,12 +689,17 @@ def _generate_response(prompt: str, app_config=None) -> str:
# 避免硬编码的通用模型 id 随权益变化而失效。
request_model = model_name or "kimi-for-coding"
logger.info(f"requesting kimi code chat completion, model: {request_model}")
+ # Provider 自身的宽松超时在连接测试场景下要收紧到测试预算以内。
+ effective_timeout = timeout_seconds
+ if request_timeout is not None:
+ effective_timeout = min(effective_timeout, request_timeout)
client = sdk_client = OpenAI(
api_key=access_token,
base_url=request_base_url,
- timeout=timeout_seconds,
+ timeout=effective_timeout,
# 服务端反爬管线要求与 CLI 一致的设备身份头。
default_headers=kimi_code_oauth.api_headers(),
+ **_sdk_transport_options(None, max_retries),
)
response = client.chat.completions.create(
model=request_model, messages=[{"role": "user", "content": prompt}]
@@ -652,6 +715,7 @@ def _generate_response(prompt: str, app_config=None) -> str:
client = sdk_client = OpenAI(
api_key=api_key,
base_url=base_url,
+ **_sdk_transport_options(request_timeout, max_retries),
)
response = sdk_stream = client.chat.completions.create(
model=model_name,
@@ -677,6 +741,7 @@ def _generate_response(prompt: str, app_config=None) -> str:
client = sdk_client = OpenAI(
api_key=api_key,
base_url=base_url,
+ **_sdk_transport_options(request_timeout, max_retries),
)
response = client.chat.completions.create(
@@ -716,10 +781,17 @@ def test_connection() -> tuple[bool, str, float]:
连接测试直接复用 `_generate_response()`,因此会覆盖 API Key、Base URL、
模型名称和 Provider 专用字段,但不会进入脚本生成的重试逻辑,也不会发送
- 用户的视频主题或文案。返回值依次为成功状态、错误信息和请求耗时。
+ 用户的视频主题或文案。请求使用有界的短超时并禁用 SDK 自动重试:端点
+ 不可达(代理、防火墙或 DNS 故障)时快速失败,并附上排查方向,而不是
+ 挂起数分钟后只留下 "Request timed out."(issue #1642)。返回值依次为
+ 成功状态、错误信息和请求耗时。
"""
started_at = perf_counter()
- response = _generate_response(prompt="Reply with exactly: OK")
+ response = _generate_response(
+ prompt="Reply with exactly: OK",
+ request_timeout=LLM_CONNECTION_TEST_TIMEOUT_SECONDS,
+ max_retries=LLM_CONNECTION_TEST_MAX_RETRIES,
+ )
elapsed = perf_counter() - started_at
if not response:
@@ -728,7 +800,9 @@ def test_connection() -> tuple[bool, str, float]:
return False, error_message, elapsed
if response.startswith("Error:"):
- error_message = response.removeprefix("Error:").strip()
+ error_message = _explain_connection_test_failure(
+ response.removeprefix("Error:").strip()
+ )
logger.warning(f"llm connection test failed: {error_message}")
return False, error_message, elapsed
diff --git a/test/services/test_llm.py b/test/services/test_llm.py
index 2e043693..1c3703ab 100644
--- a/test/services/test_llm.py
+++ b/test/services/test_llm.py
@@ -301,6 +301,13 @@ class TestScriptPromptOptions(unittest.TestCase):
class TestLLMConnection(unittest.TestCase):
+ def setUp(self):
+ self.original_app_config = dict(config.app)
+
+ def tearDown(self):
+ config.app.clear()
+ config.app.update(self.original_app_config)
+
def test_connection_sends_one_minimal_request(self):
"""连接测试只发送一次固定最小请求,不触发脚本生成重试。"""
with (
@@ -309,9 +316,60 @@ class TestLLMConnection(unittest.TestCase):
):
result = llm.test_connection()
- generate.assert_called_once_with(prompt="Reply with exactly: OK")
+ generate.assert_called_once_with(
+ prompt="Reply with exactly: OK",
+ request_timeout=llm.LLM_CONNECTION_TEST_TIMEOUT_SECONDS,
+ max_retries=llm.LLM_CONNECTION_TEST_MAX_RETRIES,
+ )
self.assertEqual(result, (True, "", 0.25))
+ def test_connection_applies_bounded_timeout_and_disables_sdk_retries(self):
+ """连接测试的短超时与禁用重试必须作用到实际 SDK 客户端(issue #1642)。"""
+ config.app.update(
+ {
+ "llm_provider": "openai",
+ "openai_api_key": "test-key",
+ "openai_model_name": "test-model",
+ "openai_base_url": "https://example.test/v1",
+ }
+ )
+ captured = {}
+
+ def create(**kwargs):
+ captured.update(kwargs)
+ return types.SimpleNamespace(
+ chat=types.SimpleNamespace(
+ completions=types.SimpleNamespace(create=lambda **kw: None)
+ )
+ )
+
+ with patch.object(llm, "OpenAI", side_effect=create):
+ llm.test_connection()
+
+ self.assertEqual(
+ captured.get("timeout"), llm.LLM_CONNECTION_TEST_TIMEOUT_SECONDS
+ )
+ self.assertEqual(
+ captured.get("max_retries"), llm.LLM_CONNECTION_TEST_MAX_RETRIES
+ )
+
+ def test_connection_failure_hint_for_timeout_errors(self):
+ """超时/连接失败要附上排查方向,而不是只显示 SDK 原始文案。"""
+ with (
+ patch.object(
+ llm, "_generate_response", return_value="Error: Request timed out."
+ ),
+ patch.object(llm, "perf_counter", side_effect=[0.0, 5.0]),
+ ):
+ ok, error, elapsed = llm.test_connection()
+
+ self.assertFalse(ok)
+ self.assertEqual(elapsed, 5.0)
+ self.assertIn("Request timed out.", error)
+ self.assertIn(f"{llm.LLM_CONNECTION_TEST_TIMEOUT_SECONDS:.0f}s", error)
+ self.assertIn("reachable", error)
+ self.assertIn("proxy", error)
+
def test_connection_returns_provider_error(self):
"""Provider 返回错误时应保留可诊断信息,并报告本次请求耗时。"""
with (
@@ -337,6 +395,106 @@ class TestLLMConnection(unittest.TestCase):
self.assertEqual(result, (False, "LLM returned an empty response", 1.0))
+class TestGenerateResponseTransportLimits(unittest.TestCase):
+ def setUp(self):
+ self.original_app_config = dict(config.app)
+
+ def tearDown(self):
+ config.app.clear()
+ config.app.update(self.original_app_config)
+
+ @staticmethod
+ def _fake_client():
+ """返回一个最小可用客户端,让 `_generate_response` 走完请求分支。"""
+ message = types.SimpleNamespace(content="ok")
+ choice = types.SimpleNamespace(message=message)
+ completion = types.SimpleNamespace(choices=[choice])
+ return types.SimpleNamespace(
+ chat=types.SimpleNamespace(
+ completions=types.SimpleNamespace(create=lambda **kwargs: completion)
+ )
+ )
+
+ def test_transport_limits_forwarded_to_openai_compatible_clients(self):
+ """显式传入的超时/重试约束必须作用到全部 OpenAI 兼容客户端。"""
+ 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:
+ with self.subTest(provider=provider):
+ config.app.update({"llm_provider": provider, **settings})
+ with (
+ patch.object(
+ llm,
+ constructor,
+ return_value=self._fake_client(),
+ ) as client_constructor,
+ patch.object(llm, "ChatCompletion", types.SimpleNamespace),
+ ):
+ result = llm._generate_response(
+ "hello",
+ request_timeout=12.5,
+ max_retries=0,
+ )
+
+ self.assertEqual(result, "ok", result)
+ kwargs = client_constructor.call_args.kwargs
+ self.assertEqual(kwargs.get("timeout"), 12.5)
+ self.assertEqual(kwargs.get("max_retries"), 0)
+
+ def test_default_generation_keeps_sdk_defaults(self):
+ """不传约束时客户端构造参数保持原样,生成链路行为不变。"""
+ config.app.update(
+ {
+ "llm_provider": "openai",
+ "openai_api_key": "test-key",
+ "openai_model_name": "test-model",
+ "openai_base_url": "https://example.test/v1",
+ }
+ )
+ with (
+ patch.object(
+ llm, "OpenAI", return_value=self._fake_client()
+ ) as client_constructor,
+ patch.object(llm, "ChatCompletion", types.SimpleNamespace),
+ ):
+ result = llm._generate_response("hello")
+
+ self.assertEqual(result, "ok", result)
+ client_constructor.assert_called_once_with(
+ api_key="test-key",
+ base_url="https://example.test/v1",
+ )
+
+
class TestLiteLLMProvider(unittest.TestCase):
def setUp(self):
self.original_app_config = dict(config.app)