fix: tolerate stream chunks without a choices key in stream_chunk_builder (#34382)

The Responses-API to /chat/completions bridge yields ModelResponseStream
chunks that carry choices followed by a trailing event object that has no
choices key. stream_chunk_builder assumed every chunk was subscriptable at
"choices", so assembling those chunks raised KeyError('choices') and was
re-wrapped as a 500 APIError building the streaming usage.

Guard each choices access with .get("choices") so choices-less chunks are
skipped instead of crashing. Behavior is unchanged for chunks that do carry
choices, since .get("choices") is truthy only for a non-empty choices list.

Adds a regression test that assembles content across chunks followed by a
trailing chunk with no choices key.

Co-authored-by: mubashir1osmani <mubashir.osmani777@gmail.com>

litellm/main.py

@@ -8568,7 +8568,7 @@ def stream_chunk_builder(
         if len(chunks) == 0:
             return None
         ## Route to the text completion logic
-        first_chunk_with_choices: Final = next((c for c in chunks if c["choices"]), None)
+        first_chunk_with_choices: Final = next((c for c in chunks if c.get("choices")), None)
         if first_chunk_with_choices is not None and isinstance(
             first_chunk_with_choices["choices"][0], litellm.utils.TextChoices
         ):  # route to the text completion logic
@@ -8583,7 +8583,7 @@ def stream_chunk_builder(
         simple_content_parts: Final[list[str]] = []
         is_simple_text_stream = True
         for chunk in chunks:
-            if len(chunk["choices"]) == 0:
+            if not chunk.get("choices"):
                 continue
 
             choice = chunk["choices"][0]
@@ -8649,7 +8649,7 @@ def stream_chunk_builder(
         tool_call_chunks: Final = [
             chunk
             for chunk in chunks
-            if len(chunk["choices"]) > 0
+            if chunk.get("choices")
             and "tool_calls" in chunk["choices"][0]["delta"]
             and chunk["choices"][0]["delta"]["tool_calls"] is not None
         ]
@@ -8663,7 +8663,7 @@ def stream_chunk_builder(
         function_call_chunks: Final = [
             chunk
             for chunk in chunks
-            if len(chunk["choices"]) > 0
+            if chunk.get("choices")
             and "function_call" in chunk["choices"][0]["delta"]
             and chunk["choices"][0]["delta"]["function_call"] is not None
         ]
@@ -8676,7 +8676,7 @@ def stream_chunk_builder(
         content_chunks: Final = [
             chunk
             for chunk in chunks
-            if len(chunk["choices"]) > 0
+            if chunk.get("choices")
             and "content" in chunk["choices"][0]["delta"]
             and chunk["choices"][0]["delta"]["content"] is not None
         ]
@@ -8687,7 +8687,7 @@ def stream_chunk_builder(
         thinking_blocks: Final = [
             chunk
             for chunk in chunks
-            if len(chunk["choices"]) > 0
+            if chunk.get("choices")
             and "thinking_blocks" in chunk["choices"][0]["delta"]
             and chunk["choices"][0]["delta"]["thinking_blocks"] is not None
         ]
@@ -8700,7 +8700,7 @@ def stream_chunk_builder(
         reasoning_chunks: Final = [
             chunk
             for chunk in chunks
-            if len(chunk["choices"]) > 0
+            if chunk.get("choices")
             and "reasoning_content" in chunk["choices"][0]["delta"]
             and chunk["choices"][0]["delta"]["reasoning_content"] is not None
         ]
@@ -8713,7 +8713,7 @@ def stream_chunk_builder(
         annotation_chunks: Final = [
             chunk
             for chunk in chunks
-            if len(chunk["choices"]) > 0
+            if chunk.get("choices")
             and "annotations" in chunk["choices"][0]["delta"]
             and chunk["choices"][0]["delta"]["annotations"] is not None
         ]
@@ -8730,7 +8730,7 @@ def stream_chunk_builder(
         audio_chunks: Final = [
             chunk
             for chunk in chunks
-            if len(chunk["choices"]) > 0
+            if chunk.get("choices")
             and "audio" in chunk["choices"][0]["delta"]
             and chunk["choices"][0]["delta"]["audio"] is not None
         ]
@@ -8744,7 +8744,7 @@ def stream_chunk_builder(
         image_chunks: Final = [
             chunk
             for chunk in chunks
-            if len(chunk["choices"]) > 0
+            if chunk.get("choices")
             and "images" in chunk["choices"][0]["delta"]
             and chunk["choices"][0]["delta"]["images"] is not None
         ]
@@ -8761,7 +8761,7 @@ def stream_chunk_builder(
         provider_specific_chunks: Final = [
             chunk
             for chunk in chunks
-            if len(chunk["choices"]) > 0
+            if chunk.get("choices")
             and "provider_specific_fields" in chunk["choices"][0]["delta"]
             and chunk["choices"][0]["delta"]["provider_specific_fields"] is not None
         ]

tests/test_litellm/litellm_core_utils/test_streaming_chunk_builder_utils.py

@@ -989,6 +989,32 @@ def test_cost_field_in_usage_chunks():
     assert usage.completion_tokens == 5
 
 
+def test_stream_chunk_builder_tolerates_trailing_chunk_without_choices():
+    """Regression for https://github.com/BerriAI/litellm/issues/32051
+
+    The Responses-API bridge yields ModelResponseStream chunks with choices
+    followed by a trailing event object that has no ``choices`` key. Building
+    those chunks used to raise ``KeyError('choices')`` (surfaced as a 500
+    APIError); it must now skip the choices-less chunk and assemble content.
+    """
+    from litellm.types.llms.base import BaseLiteLLMOpenAIResponseObject
+
+    content_chunks = [
+        ModelResponseStream(
+            model="gpt-4o",
+            choices=[StreamingChoices(index=0, delta=Delta(content=part))],
+        )
+        for part in ("Hello", " world")
+    ]
+    trailing_chunk = BaseLiteLLMOpenAIResponseObject()
+    assert "choices" not in trailing_chunk
+
+    response = stream_chunk_builder(chunks=content_chunks + [trailing_chunk])
+
+    assert response is not None
+    assert response.choices[0].message.content == "Hello world"
+
+
 def test_anthropic_speed_and_geo_survive_stream_assembly():
     """Anthropic prices fast mode and non-global regions with a multiplier read off
     ``usage.speed`` / ``usage.inference_geo``. Dropping them while reassembling a stream