Merge 2bc189ff8f into 6f5ad78a1f

litellm/llms/openai/openai.py

@@ -358,6 +358,27 @@ def _embedding_request_without_sdk_defaults(
     return body, options
 
 
+def _status_code_for_openai_sdk_error(e: Exception) -> int:
+    """
+    Status code to attribute to an exception raised by the OpenAI SDK.
+
+    A bare ``openai.OpenAIError`` is raised at client construction, before any HTTP
+    exchange, so it carries no status code; missing credentials is the common case.
+    Defaulting it to 500 asserts that a server responded with a server error, which
+    makes a permanently unfixable configuration error retryable. Subclasses such as
+    ``APIConnectionError`` and ``APIStatusError`` keep their existing treatment, so
+    genuinely transient failures are still retried.
+
+    See https://github.com/BerriAI/litellm/issues/35860
+    """
+    status_code: Final = getattr(e, "status_code", None)
+    if status_code is not None:
+        return status_code
+    if type(e) is openai.OpenAIError:
+        return 400
+    return 500
+
+
 class OpenAIChatCompletion(BaseLLM, BaseOpenAILLM):
     def __init__(self) -> None:
         super().__init__()
@@ -998,7 +1019,7 @@ class OpenAIChatCompletion(BaseLLM, BaseOpenAILLM):
                 # e.message
             except Exception as e:
                 exception_response = getattr(e, "response", None)
-                status_code = getattr(e, "status_code", 500)
+                status_code = _status_code_for_openai_sdk_error(e)
                 exception_body = getattr(e, "body", None)
                 error_headers = getattr(e, "headers", None)
                 if error_headers is None and exception_response:
@@ -1153,7 +1174,7 @@ class OpenAIChatCompletion(BaseLLM, BaseOpenAILLM):
                     raise e
 
                 error_headers = getattr(e, "headers", None)
-                status_code = getattr(e, "status_code", 500)
+                status_code = _status_code_for_openai_sdk_error(e)
                 error_response = getattr(e, "response", None)
                 exception_body = getattr(e, "body", None)
                 if error_headers is None and error_response:
@@ -1181,8 +1202,14 @@ class OpenAIChatCompletion(BaseLLM, BaseOpenAILLM):
                             body=exception_body,
                         )
                     else:
+                        # status_code was computed above by
+                        # _status_code_for_openai_sdk_error; hard-coding 500
+                        # here made the no-response branch discard it, so an
+                        # async streaming call with no configured API key
+                        # surfaced as a retryable 500 instead of the 400 the
+                        # non-streaming path reports.
                         raise OpenAIError(
-                            status_code=500,
+                            status_code=status_code,
                             message=f"{e}",
                             headers=error_headers,
                             body=exception_body,

tests/unit/llms/openai/test_openai_sdk_error_status_code.py

@@ -0,0 +1,74 @@
+import httpx
+import openai
+import pytest
+
+from litellm.llms.openai.openai import _status_code_for_openai_sdk_error
+
+
+def test_missing_credentials_is_not_a_server_error():
+    """
+    Regression test for https://github.com/BerriAI/litellm/issues/35860
+
+    The SDK raises a bare OpenAIError at client construction when no key is
+    configured. That happens before any HTTP exchange, so calling it a 500 marks a
+    permanently unfixable configuration error as a retryable server error.
+    """
+    with pytest.raises(openai.OpenAIError) as exc_info:
+        openai.OpenAI(api_key=None)
+
+    error = exc_info.value
+    assert not hasattr(error, "status_code")
+    assert _status_code_for_openai_sdk_error(error) == 400
+
+
+def test_connection_error_stays_retryable():
+    """Transient failures must keep the 500 default so retries still happen."""
+    error = openai.APIConnectionError(request=httpx.Request("POST", "http://example.com"))
+
+    assert not hasattr(error, "status_code")
+    assert _status_code_for_openai_sdk_error(error) == 500
+
+
+def test_existing_status_code_is_preserved():
+    class _WithStatus(Exception):
+        status_code = 429
+
+    assert _status_code_for_openai_sdk_error(_WithStatus()) == 429
+
+
+def test_unknown_exception_keeps_the_server_error_default():
+    """Anything that is not a bare OpenAIError keeps the previous behaviour."""
+    assert _status_code_for_openai_sdk_error(ValueError("boom")) == 500
+
+
+def test_status_code_zero_is_preserved():
+    """A falsy-but-present status code must not fall through to the default."""
+
+    class _ZeroStatus(Exception):
+        status_code = 0
+
+    assert _status_code_for_openai_sdk_error(_ZeroStatus()) == 0
+
+
+@pytest.mark.asyncio
+async def test_async_streaming_missing_credentials_is_not_a_server_error(monkeypatch):
+    """
+    The helper alone is not enough: async_streaming's no-response branch used
+    to hard-code 500, discarding the status computed above it, so the same
+    missing-key error that maps to 400 on the non-streaming path surfaced as a
+    retryable 500 when stream=True. This exercises the full call path.
+    """
+    import litellm
+
+    monkeypatch.delenv("OPENAI_API_KEY", raising=False)
+    monkeypatch.setattr(litellm, "api_key", None, raising=False)
+    monkeypatch.setattr(litellm, "openai_key", None, raising=False)
+
+    with pytest.raises(openai.OpenAIError) as exc_info:
+        await litellm.acompletion(
+            model="gpt-4o-mini",
+            messages=[{"role": "user", "content": "hi"}],
+            stream=True,
+        )
+
+    assert getattr(exc_info.value, "status_code", None) == 400