fix(convert_dict_to_response): handle empty choices list without raising 500 APIError

litellm/litellm_core_utils/llm_response_utils/convert_dict_to_response.py

@@ -3,7 +3,7 @@
 import re
 import time
 import traceback
-from collections.abc import Iterable, Sequence
+from collections.abc import Sequence
 from typing import Final, Literal, cast
 
 import litellm
@@ -179,7 +179,7 @@ async def convert_to_streaming_response_async(
 
     choice_list: Final[list[StreamingChoices]] = []
 
-    if not response_object.get("choices"):
+    if not isinstance(response_object.get("choices"), list):
         from litellm.exceptions import APIError
 
         raise APIError(
@@ -287,7 +287,7 @@ def convert_to_streaming_response(
     model_response_object: Final = ModelResponseStream()
     choice_list: Final[list[StreamingChoices]] = []
 
-    if not response_object.get("choices"):
+    if not isinstance(response_object.get("choices"), list):
         from litellm.exceptions import APIError
 
         raise APIError(
@@ -623,7 +623,7 @@ def convert_to_model_response_object(
                 return convert_to_streaming_response(response_object=response_object)
             choice_list: Final[list[Choices]] = []
 
-            if not response_object.get("choices") or not isinstance(response_object["choices"], Iterable):
+            if not isinstance(response_object.get("choices"), list):
                 from litellm.exceptions import APIError
 
                 raise APIError(

litellm/litellm_core_utils/streaming_handler.py

@@ -1473,17 +1473,14 @@ def _dispatch_provider_chunk(
                 self.received_finish_reason = response_obj["finish_reason"]
         elif self.custom_llm_provider == "cached_response":
             cached_chunk: Final = cast(ModelResponseStream, chunk)
-            chunk_finish_reason: Final = cached_chunk.choices[0].finish_reason
+            cached_choice: Final = cached_chunk.choices[0] if cached_chunk.choices else None
+            chunk_finish_reason: Final = cached_choice.finish_reason if cached_choice is not None else None
             response_obj = {
-                "text": cached_chunk.choices[0].delta.content,
+                "text": cached_choice.delta.content if cached_choice is not None else None,
                 "is_finished": chunk_finish_reason is not None,
                 "finish_reason": chunk_finish_reason,
                 "original_chunk": cached_chunk,
-                "tool_calls": (
-                    cached_chunk.choices[0].delta.tool_calls
-                    if hasattr(cached_chunk.choices[0].delta, "tool_calls")
-                    else None
-                ),
+                "tool_calls": (getattr(cached_choice.delta, "tool_calls", None) if cached_choice is not None else None),
             }
 
             completion_obj["content"] = response_obj["text"]

litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py

@@ -1487,8 +1487,9 @@ def translate_openai_response_to_anthropic(
             anthropic_content.insert(0, polyfill_result.compaction_block)
 
         ## extract finish reason
+        openai_finish_reason: Final = response.choices[0].finish_reason if response.choices else "stop"
         translated_finish_reason: Final = self._translate_openai_finish_reason_to_anthropic(
-            openai_finish_reason=response.choices[0].finish_reason
+            openai_finish_reason=openai_finish_reason
         )
         anthropic_finish_reason: Final = (
             "refusal"

tests/llm_translation/test_llm_response_utils/test_convert_dict_to_chat_completion.py

@@ -1623,29 +1623,26 @@ def test_convert_to_model_response_object_no_choices_raises_api_error(self):
 
         assert "no 'choices'" in exc_info.value.message
 
-    def test_convert_to_model_response_object_empty_choices_raises_api_error(self):
-        """Empty choices list raises APIError, same as missing/null choices.
+    def test_convert_to_model_response_object_empty_choices_returns_empty_list(self):
+        """An empty choices list is a real provider answer, so it converts to choices=[] instead of raising.
 
-        Provider-specific repair (e.g. github_copilot synthesizing choices for
-        Anthropic-native responses) happens before this guard, in the provider
-        config; the core utility keeps treating empty choices as an error.
+        See: https://github.com/BerriAI/litellm/issues/40276
         """
-        from litellm.exceptions import APIError
-
         response_object = {
             "id": "msg_123",
             "model": "some-model",
             "choices": [],
             "usage": {"prompt_tokens": 10, "completion_tokens": 1, "total_tokens": 11},
         }
 
-        with pytest.raises(APIError) as exc_info:
-            convert_to_model_response_object(
-                response_object=response_object,
-                model_response_object=ModelResponse(),
-            )
+        result = convert_to_model_response_object(
+            response_object=response_object,
+            model_response_object=ModelResponse(),
+        )
 
-        assert "no 'choices'" in exc_info.value.message
+        assert isinstance(result, ModelResponse)
+        assert result.choices == []
+        assert result.usage.prompt_tokens == 10
 
     def test_convert_to_model_response_object_null_choices_raises_api_error(self):
         """choices=None raises APIError."""

tests/test_litellm/litellm_core_utils/llm_response_utils/test_convert_dict_to_response.py

@@ -1,4 +1,6 @@
+from typing import Final
 
+import pytest
 
 from litellm.constants import RESPONSE_FORMAT_TOOL_NAME
 from litellm.litellm_core_utils.llm_response_utils.convert_dict_to_response import (
@@ -99,3 +101,96 @@ def test_handle_invalid_parallel_tool_calls_skips_custom_tool_calls():
     )
     result = _handle_invalid_parallel_tool_calls([custom_tool_call, function_tool_call])
     assert result == [custom_tool_call, function_tool_call]
+
+
+def test_convert_empty_choices_response() -> None:
+    from litellm.litellm_core_utils.llm_response_utils.convert_dict_to_response import (
+        convert_to_streaming_response,
+    )
+
+    resp: Final = {
+        "id": "x",
+        "created": 1,
+        "model": "gemini-3.5-flash",
+        "object": "chat.completion",
+        "choices": [],
+        "usage": {"prompt_tokens": 10, "completion_tokens": 0, "total_tokens": 10},
+        "vertex_ai_safety_results": ["blocked"],
+    }
+    result: Final = convert_to_model_response_object(
+        response_object=resp,
+        model_response_object=ModelResponse(),
+        response_type="completion",
+    )
+    assert result.choices == []
+    assert getattr(result, "vertex_ai_safety_results") == ["blocked"]
+
+    sync_stream: Final = list(convert_to_streaming_response(response_object=resp))
+    assert len(sync_stream) == 1
+    assert sync_stream[0].choices == []
+
+
+@pytest.mark.asyncio
+async def test_convert_empty_choices_response_async() -> None:
+    from litellm.litellm_core_utils.llm_response_utils.convert_dict_to_response import (
+        convert_to_streaming_response_async,
+    )
+
+    resp: Final = {
+        "id": "x",
+        "created": 1,
+        "model": "gemini-3.5-flash",
+        "object": "chat.completion",
+        "choices": [],
+        "usage": {"prompt_tokens": 10, "completion_tokens": 0, "total_tokens": 10},
+    }
+    async_chunks: Final = [chunk async for chunk in convert_to_streaming_response_async(response_object=resp)]
+    assert len(async_chunks) == 1
+    assert async_chunks[0].choices == []
+
+
+def test_convert_missing_choices_raises_api_error() -> None:
+    from litellm.exceptions import APIError
+
+    resp: Final = {
+        "id": "x",
+        "created": 1,
+        "model": "gemini-3.5-flash",
+        "object": "chat.completion",
+    }
+    with pytest.raises(APIError) as exc_info:
+        convert_to_model_response_object(
+            response_object=resp,
+            model_response_object=ModelResponse(),
+            response_type="completion",
+        )
+    assert "no 'choices'" in str(exc_info.value)
+
+
+@pytest.mark.parametrize("choices", [{}, "", None, 0])
+@pytest.mark.asyncio
+async def test_convert_non_list_choices_raises_api_error(choices: object) -> None:
+    from litellm.exceptions import APIError
+    from litellm.litellm_core_utils.llm_response_utils.convert_dict_to_response import (
+        convert_to_streaming_response,
+        convert_to_streaming_response_async,
+    )
+
+    resp: Final = {
+        "id": "x",
+        "created": 1,
+        "model": "gemini-3.5-flash",
+        "object": "chat.completion",
+        "choices": choices,
+    }
+    with pytest.raises(APIError, match="no 'choices'"):
+        convert_to_model_response_object(
+            response_object=resp,
+            model_response_object=ModelResponse(),
+            response_type="completion",
+        )
+    with pytest.raises(APIError, match="no 'choices'"):
+        list(convert_to_streaming_response(response_object=resp))
+    with pytest.raises(APIError, match="no 'choices'"):
+        async for _ in convert_to_streaming_response_async(response_object=resp):
+            pass

tests/test_litellm/litellm_core_utils/test_streaming_handler.py

@@ -6,7 +6,7 @@
 
 import asyncio
 import traceback
-from typing import Optional
+from typing import Final, Optional
 
 import litellm
 from litellm import verbose_logger
@@ -2633,6 +2633,47 @@ def test_dispatch_cached_response_extracts_delta(
     assert initialized_custom_stream_wrapper.response_id == "chatcmpl-cache-1"
 
 
+def test_dispatch_cached_response_without_choices_is_an_empty_chunk(
+    initialized_custom_stream_wrapper: CustomStreamWrapper,
+):
+    """A cached completion with no choices replays as an empty, unfinished chunk
+    instead of raising IndexError on choices[0]."""
+    initialized_custom_stream_wrapper.custom_llm_provider = "cached_response"
+    chunk: Final = ModelResponseStream(id="chatcmpl-cache-empty", choices=[])
+
+    result, model_response, completion_obj = _run_dispatch(
+        initialized_custom_stream_wrapper, chunk
+    )
+
+    assert isinstance(result, _ProviderChunkParsed)
+    assert completion_obj["content"] is None
+    assert initialized_custom_stream_wrapper.received_finish_reason is None
+    assert model_response.id == "chatcmpl-cache-empty"
+
+
+@pytest.mark.asyncio
+async def test_cached_response_without_choices_streams_a_single_stop_chunk(
+    logging_obj: Logging,
+):
+    """A stream cache hit on a completion stored with choices == [] ends with one
+    finish_reason=stop chunk, the same shape the live empty stream produced."""
+
+    async def cached_chunks():
+        yield ModelResponseStream(id="chatcmpl-cache-empty", choices=[])
+
+    wrapper: Final = CustomStreamWrapper(
+        completion_stream=cached_chunks(),
+        model="test-model",
+        logging_obj=logging_obj,
+        custom_llm_provider="cached_response",
+    )
+
+    chunks: Final = tuple([chunk async for chunk in wrapper])
+
+    assert tuple(choice.finish_reason for chunk in chunks for choice in chunk.choices) == ("stop",)
+    assert all(choice.delta.content in (None, "") for chunk in chunks for choice in chunk.choices)
+
+
 def test_dispatch_vertex_ai_legacy_text_and_finish_reason(
     initialized_custom_stream_wrapper: CustomStreamWrapper,
 ):

tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py

@@ -41,6 +41,21 @@
 )
 
 
+def test_translate_openai_response_to_anthropic_empty_choices() -> None:
+    response: Final = ModelResponse(
+        id="chatcmpl-empty",
+        model="gemini-3.5-flash",
+        choices=[],
+        usage=Usage(prompt_tokens=10, completion_tokens=0, total_tokens=10),
+    )
+
+    result: Final = LiteLLMAnthropicMessagesAdapter().translate_openai_response_to_anthropic(response)
+
+    assert result["content"] == []
+    assert result["stop_reason"] == "end_turn"
+    assert result["usage"]["input_tokens"] == 10
+
+
 def test_translate_chat_refusal_to_anthropic_response():
     response = ModelResponse(
         id="chatcmpl-refusal",

tests/test_litellm/test_main.py

@@ -120,7 +120,7 @@ def test_completion_missing_role(openai_api_response):
     print(f"openai_api_response: {openai_api_response}")
 
     with patch.object(
-        client.chat.completions.with_raw_response, "create", mock_raw_response
+        client.chat.completions.with_raw_response, "create", MagicMock(return_value=mock_raw_response)
     ) as mock_create:
         litellm.completion(
             model="gpt-4o-mini",